Archive fork in favor of upstream and DUMAS nodes
tests / test (push) Canceled after 0s
tests / windows-storage (push) Canceled after 0s

This commit is contained in:
2026-10-20 10:46:16 +00:00
parent 881dcbcfba
commit 3716b02a57
76 changed files with 12140 additions and 4410 deletions
+66
View File
@@ -0,0 +1,66 @@
# Issue #31 — Attention and Spectrum reuse signatures (3.9.5)
## 3.9.10 — Spectrum audit portability
The repair treats source CRLF and LF line endings as equivalent
in both the provider-file audit and provider module closures. All other source
bytes and the exact supported Python file sets remain checked. The canonical
source hashes are unchanged, so existing valid Python 3.13 signatures retain
their identities.
Separate `spectrum_provider_v1_py312.json` and
`spectrum_provider_v029_py312.json` profiles identify the same audited source sets
under Python 3.12. They were generated and checked with Python 3.12.10; the
original profiles remain byte-identical for Python 3.13.12. Other Python runtimes,
altered implementations, foreign closures and unknown runtime fields remain
non-reusable. The adapter does not execute provider functions to approve reuse.
This repairs Spectrum observation, not every Core/Kitchen/KJ combination.
Independent Attention/provider checks must also pass. In particular, an older
Core/Kitchen environment can still fail its own Attention observation even when
Spectrum is recognized. Do not upgrade Python or replace Attention solely to
make an old Take reusable. Keep old history and start a new Run after updating.
No new GPU or reporter-exact acceptance is claimed for this repair.
The standard CPU/JS gate passed **2,096 tests / 2 skips / 2 known expected
Inspector failures**, with compile and offline registration checks passing.
Separate-process storage probes passed 16 Python 3.13 PyTorch/Kitchen cases and
eight Python 3.12 Spectrum/default-model cases; actual Take read-only JavaScript
matched all 24. The six previously valid full Sampling Contracts stayed exact.
These CPU checks execute no Attention kernel or H3 checkpoint inference.
### 日本語 — 3.9.10の局所修正
Spectrum監査でCRLF/LFの改行差を同一として扱い、対応済みソースにPython 3.12用の
固定プロファイルを追加します。既存3.13の正常な識別値と旧Takeは維持します。
未知のコード・追加状態の拒否は継続します。Spectrumが識別可能でも、Core/Kitchenなど
別の観測に失敗する環境は再利用不可のままです。適用後は新規Runで確認してください。
旧履歴を利用可能にする目的でPythonやAttentionを変更する必要はありません。
## 3.9.8 Preview Override Phase A (2026-10-16)
The new adapter is restricted to the KJ source profile in `preview_override_provider_v1.json`: Preview Override SHA-256 `d07c8e5ff7d7abd3e31acefe2cf29f618b5da5341b3b434e72ec2ac0953230b0`, audited sibling helpers and Core preview/dispatcher files. The complete profile digest is part of the persistent identity, so later dependency-profile upgrades cannot retain the same identity accidentally. It compares loaded functions, nested code, defaults and supported classes with the pinned source; it never imports KJNodes, executes a preview callback, loads a VAE or initializes CUDA during signature observation. Source line endings are normalized to LF for provider auditing.
After that audit, the exact builtin `frames` list is treated as execution-local display data. Max resolution, node ID, JPEG quality, suppression, frame count, FPS and supported Core preview settings remain part of the identity. Core's cross-key insertion order and each wrapper list's order are preserved explicitly, because the generic dictionary observer sorts keys. The ordinary non-KJ signature path is unchanged.
Phase A requires `vae is None`, `audio_vae is None`, KJ Tiny VAE `none`, and Core NoPreviews/Auto/Latent2RGB. Unknown files, replaced code/defaults/helpers, extra wrapper state, foreign classes and approximate-VAE previews keep reuse disabled; they are not promoted by an empty frame list. This profile does not claim support for every KJNodes 1.5.2 checkout or the different upstream Preview Override source. No general `frames` exclusion is added.
Existing normal non-KJ Sampling v6 identities remain exact. Earlier safe KJ identities may differ and cannot be assumed reusable; earlier unsafe/v5 histories remain preserved and read-only. Start a new Run after updating; old KJ history is not migrated. Full real-Core CPU gate: 1,995 passed / 3 skipped / 2 existing Inspector expected failures / zero unexpected failures and CUDA initialization attempts; the 48 dedicated KJ tests passed without skips. Normal ComfyUI GPU acceptance passed four Core-API Queues: PyTorch and Comfy Kitchen INT8 + the audited local Spectrum implementation, each with separate-Queue Chunk 1 to 2 continuation and held preview images. Both second Queues reused one chunk and generated one; original Chunk 1 bytes and contracts were unchanged. All four video/audio outputs fully decoded. See docs/KJ_PREVIEW_OVERRIDE_V398.md for exact support and limits. Product files used in this targeted GPU gate are byte-identical to the 3.9.8 repair. Browser Run-button behavior, subjective preview quality, reporter-exact reproduction and per-kernel return tracing are not newly accepted.
## Previous 3.9.5 Attention / Spectrum repair
With Model Attention Backend between a LoRA Loader and Spectrum, cyclic Attention closures and a slots-based SpectrumH3Binding could be unobservable. Generation still succeeded, but Sampling v6 correctly refused reuse with "Saved history is read-only". This update identifies the reviewed runtime implementations rather than forcing resume_safe.
Core PyTorch and Comfy Kitchen INT8 overrides are identified by their loaded functions, defaults, closures, both container paths and numerical/provider state. Spectrum identity includes all 27 effective immutable config fields and audited source/loaded-function contracts. Reviewed history, counters and statistics do not change persistent identity. Unknown sources, foreign module closures, extra runtime state and altered implementations remain non-reusable.
Two explicit Spectrum source profiles are included: upstream v0.2.29 at commit 806fe498d574a726fecb8818917f00a4b993fd77 (55 provider files), and the prior audited 44-file implementation. Versions/class names alone do not enable reuse. The loaded-function contract is environment-specific: validation uses Python 3.13.12, ComfyUI Core 0.40.0 and comfy-kitchen 0.2.37; other source/runtime combinations may remain read-only. BSA/Flow/Untwist and KJ/Sage patch-node combinations are outside this targeted acceptance.
Update the custom node and restart ComfyUI. Start a new Run. Existing unsafe/v5 Runs are preserved and are not repaired or promoted; normal unaffected Sampling v6 identities remain compatible. Sampling v6, Graph v4, Run Storage v3, public IDs/socket order and shipped workflow bytes stay unchanged.
The signature gate passed 152 CPU tests with one original-profile-only skip and no failures; CUDA initialization attempts were zero. Full standard validation passed 1,893 tests with 2 skips and 2 existing Inspector expected failures (zero unexpected failures and CUDA initialization attempts). Public upstream 0.2.29 PyTorch/Spectrum and Kitchen/Spectrum new-Run GPU gates both passed separate-Queue Chunk 1→2 continuation, prefix/contract preservation and complete 512x512 / 24fps / 10-second AV decoding. Actual CUDA backend returns were observed. Existing Inspector reconfigure expected failures, browser acceptance, visual/audio quality and benchmarks remain separate.
## 日本語
Model Attention BackendとSpectrumを併用したReview継続で、Attentionの循環参照とslots付きBindingを観測できず、生成後のTakeが再利用不可になる問題へ対応します。全27設定値・実ソース・読み込まれた関数を識別し、実行中の履歴や統計値は永続署名へ含めません。resume_safeを強制せず、未知の実装や追加状態は安全側で拒否します。
公式Spectrum v0.2.29の固定コミットと、以前に監査した44ファイルの実装を別プロファイルとして扱います。別のSpectrum版、別のPython/Core実装、BSA/Flow/UntwistやKJパッチの組合せ全般を保証するものではありません。更新・再起動後は新しいRunを作成してください。以前のunsafe履歴を自動修復せず、既存TakeとWorkflowは保持します。
+95
View File
@@ -0,0 +1,95 @@
# KJ Model Preview Override continuation — 3.9.8 / Phase A
The audited Preview Override wrapper can keep its transient image list without changing the MODEL sampling identity. The adapter retains fixed settings, source and loaded implementation checks, the complete audit-profile digest, and Core's cross-key/list wrapper execution order. It does not add a generic exception for any field named `frames`.
## 3.9.9 public-provider repair
The source checkout additionally accepts the pinned upstream implementation from
[KJNodes commit 17195fc](https://github.com/kijai/ComfyUI-KJNodes/blob/17195fcfa8a05f601b766a0f09c9733284f47416/nodes/preview_override_node.py).
Its LF-normalized Preview Override SHA-256 is
`1916295463eb790d3e8ea2b0dae2b25344549479b92e90a1d238acddb2dde3ab`,
profile `kj-preview-upstream-17195fc-phase-a` in
`preview_override_provider_public_v1.json` (canonical profile SHA-256
`da128a4324d8e0a63d893501f8d65652d25da60690afaf1447d80f2a2b67617f`). Its sibling and Core hashes are the
same as the original table below. The wrapper implementation is unchanged; the
node factory additionally resolves display IDs through Core's DynamicPrompt.
Both source and loaded-code audits remain required. This does not cover arbitrary
KJNodes versions, future Core changes or preview-VAE configurations.
The original profile JSON and successful descriptors are preserved. Profile
registration alone cannot invalidate another profile's Takes. KJ-free contracts
are unchanged. Missing or malformed optional public-profile data disables only
that profile. Ordinary generation keeps its existing fallback.
Review and Render History distinguish v5 history from v6 Takes whose generation
conditions could not be verified. Bounded saved reasons identify unsupported
preview settings or implementations; missing metadata points to Detailed Report.
Reasons come from the selected Take's storage revision and do not affect its
sampling identity or enable disabled actions.
For an unsupported KJ/Core implementation, bypass **Model Preview Override** in a
workflow copy and test a new Run. For unsupported preview VAE settings, disconnect
only that preview node's optional VAE/Audio VAE and use Tiny VAE `none`; keep the
generation VAEs connected. For Core approximate-VAE preview, use Auto or
Latent2RGB. Correct the configuration before starting a new Run: repeatedly
restarting Chunk 1 with the same unsupported setup will not repair it.
Old unsafe/v5 Takes remain readable and are never promoted or rewritten.
Full real-Core CPU regression: **2,075 passed, 2 skipped, 2 known Inspector
expected failures, zero unexpected failures**. The 90-case production-JS fixture
also passed with eight real Run Storage project vectors. These checks overlap
the pytest total and must not be added together. Unmodified 3.9.8 fixtures confirm
exact legacy/KJ-free identities and byte-identical saved Takes; all 21 existing
workflow JSON/ZIP files are unchanged. Offline registration retains twelve nodes
and passes native PackedLayout / Fixed 3x5s checks. Recorded CPU drivers attempted
zero CUDA initializations. The skips cover a local-only Spectrum descriptor
fixture on upstream Spectrum and an unavailable historical V3.5 workflow.
The additional public profile passed two targeted H3 GPU Queues with cached
preview images and unchanged Chunk 1. Normal-ComfyUI browser Continue passed
separately with the retained original local profile. See
[3.9.9 validation and limits](RELEASE_NOTES_V399.md). The historical v3.9.8 tag
and assets retain their original local-only profile.
## Exact supported profile
The published v3.9.8 package accepts only the following explicitly audited source combination. KJNodes reports 1.5.2, but that version alone does not identify the supported implementation: this Preview Override file differs from upstream. The 3.9.9 extension is described above. Files are LF-normalized before source hashing. An unsupported provider remains non-reusable rather than blocking ordinary generation.
| Source | SHA-256 |
| --- | --- |
| KJ `nodes/preview_override_node.py` | `d07c8e5ff7d7abd3e31acefe2cf29f618b5da5341b3b434e72ec2ac0953230b0` |
| KJ `nodes/tiny_vae.py` | `61bf4daa634ef6bb901ff8630071caed12e733d17e7c4ba50082dc0aa5506591` |
| KJ `nodes/ltxv_nodes.py` | `11286583c56e653dbd57f5e8eda6902c84b759f16b359b3bf0c895b43e6b5364` |
| Core `latent_preview.py` | `b323be881ade3388254a72f5463e329f6d0a263b0078b69ae4256eb4e22dbba2` |
| Core `comfy/patcher_extension.py` | `1c7b3b855278d0b95d1c1a705fa0b3c91a265f320e2f97bdeb9d15469d6d4ac7` |
Profile: `kj-preview-1.5.2-local-d07c8e5-phase-a`. Canonical full-profile SHA-256: `d146bd89895fef5a8d576134d519e63bbf4ac1c35188f70b4de4e4ebfb498fc7`. The provider JSON is shipped with the runtime adapter.
Preview Override's optional VAE and Audio VAE must be disconnected, Tiny VAE must be exactly `none`, and Core preview must be NoPreviews, Auto or Latent2RGB. A generation Video/Audio VAE is still used normally; the restriction concerns the preview node's optional connections. Core approximate-VAE/TAESD, altered helpers, unknown wrappers, extra state and other providers are outside this profile. Source alone is insufficient: loaded functions, nested code, defaults, classes and selected bindings must match the audited implementation.
## Update and history protection
Restart ComfyUI, refresh the browser and create a new Run. Do not delete or rewrite old Runs/Takes. Earlier safe KJ identities may differ; old unsafe/v5 history is not repaired or promoted. Ordinary non-KJ Sampling v6 signatures remain exact. Sampling v6, Graph v4, Run Storage v3, State/Session formats, public node inputs and official workflow files are unchanged.
## Validation
Full real-Core CPU validation: 1,995 passed, 3 skipped, 2 existing Inspector expected failures, zero unexpected failures or CUDA initialization attempts. The dedicated real-provider tests passed all 48 cases, including 0/1/multiple held frames, reconstruction, fixed-setting changes, wrapper ordering, source/code/default/helper changes, VAE exclusion, Attention/Spectrum combinations, prefix preservation and old unsafe history. Offline registration retains twelve nodes.
The initial acceptance used the same repair product bytes. A subsequent recheck installed all 464 files from the published v3.9.8 tag (`58fbd855`), restarted normal ComfyUI, verified the live 3.9.8 startup banner and saved adapter/profile identity, and passed four new-Run GPU Queues:
| New Run | First Queue | Separate second Queue |
| --- | --- | --- |
| Model Attention Backend PyTorch + KJ Preview Override | Generate Chunk 1; 8 held images | Reuse Chunk 1, generate Chunk 2; 16 held images |
| Model Attention Backend Comfy Kitchen INT8 + audited local Spectrum + KJ Preview Override | Generate Chunk 1; 8 held images | Reuse Chunk 1, generate Chunk 2; 16 held images |
In both second Queues the KJ MODEL node was cached, the Sampler executed, the stored contract/revision stayed equal, and original Chunk 1 latent bytes were unchanged. All four video/audio outputs fully decoded (first Queue 5 seconds / 120 frames; second Queue 10 seconds / 240 frames, 512x512, 24fps). Saved contracts recorded Sampling v6, resume_safe=true and the pinned adapter/profile identity. Runtime source hashes and stored descriptors verified the repair was actually loaded.
Conditions: RTX 5060 Ti 16 GB, Python 3.13.12, PyTorch 2.14.1+cu130, Core 0.41.0, frontend 1.55.17, comfy-kitchen 0.2.37; 8-step Turbo LoRA, Euler/simple, 2x5-second chunks, Review Each Chunk / Continue, saving and Auto Resume on. Preview Override: resolution 512, quality 80, one frame, 12fps, suppress default on, VAE/Audio VAE absent and Tiny VAE none; Core Auto. Spectrum used the previously audited local implementation with offline smoothing/replay, system-RAM history/archive and model-aware off. This does not newly validate upstream Spectrum 0.2.29 combined with KJ, other wrapper orders, Sage patch nodes, or all preview configurations.
These were controlled submissions through the normal Core API. Browser Run-button/save-reload behavior, reporter-exact reproduction, per-kernel CUDA return tracing, subjective audiovisual/preview quality and performance benchmarks are not newly accepted. Held PNGs were 32x32, so the image counts do not establish high-resolution preview quality. The two known Inspector reconfigure expected failures remain. Existing workflow files and prior Takes were hash-protected; no history migration occurred. GitHub Release/package verification, Registry upload/moderation and real Manager installation remain distinct publication checks.
## 日本語
監査済みKJ Preview Overrideの一時画像だけを永続署名から分離し、固定設定・実装識別・ラッパー順序を保持します。KJNodes 1.5.2全体ではなく、上表のソースと読み込み済み関数を照合できる実装だけが対象です。VAE/Audio VAE未接続・Tiny VAE none・Core NoPreviews/Auto/Latent2RGBが条件です。
更新後は再起動し、新Runを使用してください。旧Take・Workflowの削除や移行は不要で、旧unsafe履歴は再利用不可のまま保持します。CPU 1,995成功(専用48件を含む)、PyTorchおよびKitchen+ローカルSpectrumの各2 Queueで画像保持後のChunk 1→2継続、元Chunk 1保持、4本の映像・音声全デコードを確認しました。初期の同一製品コードによる受入後、公開タグv3.9.8(58fbd855)を実環境へ反映・再起動して、新Runで同じ2構成の4 Queueも再確認済みです。画質・速度・実ブラウザー操作・全KJ構成への合格は主張しません。
+905
View File
@@ -0,0 +1,905 @@
# H3 Continuum — README history and implementation reference
[Current README](../README.md) · [Illustrated user guide](USER_GUIDE.md) · [CHANGELOG](../CHANGELOG.md)
This reference preserves the release descriptions, earlier screenshots, detailed
validation results and pre-V3.8 implementation notes previously shown in README.
It was extracted from source commit `894869d9` (package 3.9.8).
Statements such as "current on main", historical installation commands, node lists
and old acceptance results retain their original context; they are not current
installation recommendations or broader validation claims.
For current downloads, compatibility and known limits, start with the README.
- [v3.9.8 and published-tag GPU conditions](#398--kj-model-preview-override-continuation)
- [Older implementation notes](#archived-implementation-notes-pre-v38-not-the-current-guide)
- [Prompt formats and examples](#prompt-formats)
- [Historical resource and validation limits](#limits)
<!-- Preserved README sections begin. -->
**Latest package: 3.9.8.** Use the fixed-version GitHub downloads below. Registry publication is verified separately.
| Version | Release | Custom-node ZIP | SHA-256 checksum |
| --- | --- | --- | --- |
| 3.9.8 | [Release](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/tag/v3.9.8) | [ZIP](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/download/v3.9.8/ComfyUI-H3-Continuum-v3.9.8.zip) | [SHA256SUMS.txt](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/download/v3.9.8/SHA256SUMS.txt) |
| 3.9.7 | [Release](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/tag/v3.9.7) | [ZIP](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/download/v3.9.7/ComfyUI-H3-Continuum-v3.9.7.zip) | [SHA256SUMS.txt](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/download/v3.9.7/SHA256SUMS.txt) |
| 3.9.6 | [Release](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/tag/v3.9.6) | [ZIP](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/download/v3.9.6/ComfyUI-H3-Continuum-v3.9.6.zip) | [SHA256SUMS.txt](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/download/v3.9.6/SHA256SUMS.txt) |
| 3.9.5 | [Release](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/tag/v3.9.5) | [ZIP](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/download/v3.9.5/ComfyUI-H3-Continuum-v3.9.5.zip) | [SHA256SUMS.txt](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/download/v3.9.5/SHA256SUMS.txt) |
| 3.9.4 | [Release](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/tag/v3.9.4) | [ZIP](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/download/v3.9.4/ComfyUI-H3-Continuum-v3.9.4.zip) | [SHA256SUMS.txt](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/download/v3.9.4/SHA256SUMS.txt) |
| 3.9.3 | [Release](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/tag/v3.9.3) | [ZIP](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/download/v3.9.3/ComfyUI-H3-Continuum-v3.9.3.zip) | [SHA256SUMS.txt](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/download/v3.9.3/SHA256SUMS.txt) |
| 3.9.2 | [Release](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/tag/v3.9.2) | [ZIP](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/download/v3.9.2/ComfyUI-H3-Continuum-v3.9.2.zip) | [SHA256SUMS.txt](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/download/v3.9.2/SHA256SUMS.txt) |
## 3.9.8 — KJ Model Preview Override continuation
An audited KJ Model Preview Override can keep preview images without changing a new Run's sampling identity. Separate Queues can reuse Chunk 1 and generate Chunk 2 while preserving earlier Takes. Support is limited to the explicitly pinned KJ source and Core helpers, **VAE / Audio VAE disconnected, Tiny VAE `none`**, and Core NoPreviews/Auto/Latent2RGB. Other KJ builds remain conservative; version 1.5.2 alone is insufficient. Restart ComfyUI, refresh the browser and start a new Run; do not delete old Takes or workflows. See [compatibility, source hashes and verified scope](KJ_PREVIEW_OVERRIDE_V398.md).
### Published v3.9.8 GPU recheck — 2026-10-16: PASS
After restarting the normal ComfyUI installation with the published tag **`v3.9.8` / `58fbd855`**, all 464 installed release files matched the tag. The startup log confirmed 3.9.8, and new saved Sampling v6 contracts contained the KJ adapter and pinned audit-profile digest. Two brand-new Runs passed four Core-API GPU Queues:
| KJ Preview Override configuration | Separate-Queue Chunk 1 → 2 |
| --- | --- |
| Model Attention Backend: PyTorch, Spectrum OFF | **PASS — 1 reused / 1 generated** |
| Model Attention Backend: Comfy Kitchen INT8 + audited local Spectrum | **PASS — 1 reused / 1 generated** |
On both second Queues, the KJ MODEL stayed cached with preview images retained (**8 → 16**), while the Sampler executed. The contract/revision and original Chunk 1 latent hashes stayed unchanged; `resume_safe=true`. All four video/audio outputs fully decoded (5 seconds / 120 frames, then 10 seconds / 240 frames, 512×512 at 24fps). Existing workflows and Takes were preserved: 170 protected files, including 71 saved workflow files and 46 pre-existing Run/Take files, retained their hashes.
**Scope:** only the exact audited KJ/Core profile; the local KJ source differs from upstream, so KJNodes 1.5.2 alone is insufficient. Preview Override VAE / Audio VAE were disconnected, Tiny VAE was `none`, Core preview was Auto. GPU: RTX 5060 Ti 16 GB, 8-step Turbo, Euler/simple, 2×5-second chunks. Spectrum used the audited local offline-replay implementation, system-RAM history/archive and model-aware OFF. **Unverified:** browser Run-button/save-reload behavior, the reporter's exact complete graph, other KJ/Spectrum builds, SolAttn/preview VAE paths, per-kernel return tracing, subjective quality and speed/VRAM benchmarks. Retained PNGs were 32×32, not high-resolution preview-quality evidence. The known Inspector issue remains; old unsafe/v5 Takes are not repaired or promoted. [Issue #31 result](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/issues/31#issuecomment-6209571261).
## 3.9.7 — External-input regeneration cache fix
Fix automatic Full Run variations when Chunks or Variation Nonce come from external nodes. Unresolved linked inputs no longer let Core skip an explicitly requested regeneration. Earlier accepted chunks, existing Takes, public inputs and official workflows are preserved. Update the custom nodes, restart ComfyUI and refresh the browser; no workflow changes or history deletion are required. [Release notes](RELEASE_NOTES_V397.md).
**Published v3.9.7 GPU recheck (2026-10-15): PASS.** After restarting with the published tag (`18df80d`), external Core INT nodes supplied Chunks=2 and Nonce=0. Two completely identical Chunk 2 regeneration prompts advanced automatic Nonce **1 → 2**; both reused Chunk 1 and generated only Chunk 2. Earlier Takes and raw latents were retained. All three outputs (baseline plus two variations) fully decoded with 240 frames and 10-second video/audio. Existing 60 workflows, one settings file, 30 Run/Take files and all 459 release files remained hash-identical. Scope: RTX 5060 Ti 16 GB, 512×512, 24 fps, 2×5 seconds, 8-step Turbo, PyTorch attention, Spectrum OFF, Core API submissions. No CUDA/OOM/execution errors were observed; this does not extend browser, subjective-quality or other-accelerator acceptance.
The original [3.9.2 GitHub source archive](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/archive/refs/tags/v3.9.2.zip) remains available from its unchanged tag. This custom-node package ZIP is distinct from the official workflow ZIPs.
<img width="1026" height="756" alt="{56C36D94-39EC-4484-9495-5648A1F486F3}" src="https://github.com/user-attachments/assets/ad9bd73f-1ac2-4925-80dd-55a5d3231431" />
## 3.9.6 — regeneration selections and automatic variations
Explicit **Regenerate From** and **Variation Nonce** survive V3.8/V3.9 workflow save/reload. With progress saving enabled, **Full Run + an explicit Chunk + Nonce 0** reaches Run Storage on every Queue, allowing another automatic variation after a completed Take. Earlier accepted chunks and previous Takes are kept. Restart ComfyUI and refresh the browser. For an older unmarked Chunk 1 / nonce 0, select Chunk 1 again and save once. See [operation, compatibility and verified scope](REGENERATE_INTENT_V396.md).
## 3.9.5 — Issue #31: reusable Attention and Spectrum contracts
Review Each Chunk can resume new Runs using Core Model Attention Backend with PyTorch or Comfy Kitchen INT8 and an explicitly audited Spectrum binding, including upstream Spectrum v0.2.29. Existing unsafe history remains readable and read-only for reuse; start a new Run after updating. Public node IDs, workflows and Sampling v6 are unchanged. See [supported implementations, update steps and validation limits](ISSUE31_SAMPLING_SIGNATURES.md).
## 3.9.4 — Registry packaging repair (2026-10-14)
Include the declared Review restoration guide and exclude repository-only prompt-authoring files from the Registry ZIP. The publishing workflow now builds the actual package and checks its exact file set, CRC and every Registry Manifest hash before uploading. Runtime behavior, twelve public node IDs, workflows and saved Runs/Takes are unchanged. GitHub Releases 3.9.3 and 3.9.2 remain available.
## 3.9.3 — choose temporary storage (2026-10-13)
Choose where Continuum writes its temporary decoded frames and Finalize buffer. Open **Settings → H3Continuum → Storage → Manage temporary storage**. Enter an absolute folder on the **ComfyUI server**, or choose a server storage candidate, then validate and save. Leave it blank to keep the existing locations. No workflow changes are required; continue using the V3.9 nodes and your 3.9.1 / 3.9.2 workflows.
- The selected folder applies to **both Finalize and Decode Cache Helper**. It contains a dedicated `h3-continuum-scratch` directory. Saved Runs/Takes and finished videos keep their existing locations.
- The panel shows current and previous locations, temporary file sizes, free space, and files in use or eligible for deletion. **Delete unused temporary files** protects live outputs, Tensor views, active processes and files whose ownership cannot be verified. Deleting a cache can require decoding and writing it again.
- Save or clean only when the running and pending queues are empty. New files use the new location; existing outputs remain valid at their old locations. A missing selected drive does not silently redirect writes to your original SSD. Shared storage controls are read-only in multi-user mode.
- Mode selection stays in the existing nodes: **Finalize: Auto / RAM / Disk-backed**; **Decode Cache Helper: Auto / RAM / Off**. Auto keeps its conservative 4 GiB final-IMAGE threshold and RAM reserve. Manual RAM now checks available memory before allocation; manual Disk-backed checks free space. These checks cannot guarantee against OOM, changing disk usage or a drive disconnect during execution. No new VRAM storage mode is added.
**Moving storage changes where writes occur, not their total volume.** With only one physical drive, moving to another folder or partition on that drive does not divert its writes. A separate SSD is usually the faster alternative; an HDD can be used but may slow decoding, assembly and saving. If sufficient CPU RAM is available, RAM modes can avoid these scratch writes, but memory pressure can cause paging or OOM. Turning Decode Cache Off does not turn off Finalize's Disk-backed buffer. File-size displays are logical sizes, not physical disk allocation or cumulative SSD writes. Local filesystems are supported; network shares are outside this feature's supported scope.
Update the custom node package, restart ComfyUI to load the backend, and hard-refresh the browser. Later changes to this storage setting take effect without a restart. Blank settings retain the ComfyUI temp directory for Finalize and `H3_DECODE_CACHE_TEMP_DIR` (or OS temp) for Decode Cache. A selected folder takes precedence over that cache environment variable.
The storage changes preserve frame dtype, video/audio assembly and generation identity. Validation and remaining platform limits are recorded in [PACKAGE_VALIDATION.txt](../PACKAGE_VALIDATION.txt). The implemented settings screenshot appears at the top of this README. GitHub Release and Registry publication status are listed above; neither follows automatically from a source update.
If the panel reports HTTP 404, inspect the ComfyUI startup log. An old process can keep the chosen address/port occupied, so the updated process fails to start while its new JavaScript is already visible. Stop the obsolete instance or choose another port, then start the updated backend. **Refresh** retrieves storage status; it does not restart ComfyUI. Failed reads now show the HTTP result and keep mutations disabled until status is available.
## 3.9.2 — Issue #30 frontend hotfix (2026-10-13)
This hotfix reduces unnecessary Review and Reference Images UI work:
- Skip upstream Review settings snapshots when no comparison baseline exists.
- Notify connected Samplers once per Reference selection, preserving immediate native values and the original widget callback.
- Skip Reference node resizing when its target dimensions already match its actual dimensions.
Review monitoring with a baseline and fresh Take/Queue checks remain in place. Sampling, saved Runs/Takes, public node IDs and official workflow bytes are unchanged. Keep using the V3.9 nodes.
The frontend repair passed **1,703 CPU/JavaScript regression tests, 2 skipped**, with **2 explicit expected failures** for the independent Reference Plan Inspector bug after repeated Sampler configuration. Offline Core registration/planning and isolated Chromium checks passed. Real ComfyUI browser checks for this package are recorded in [PACKAGE_VALIDATION.txt](../PACKAGE_VALIDATION.txt); mobile freeze resolution is unconfirmed. Reference table rebuilding and Review snapshots with a baseline remain. No new GPU generation is part of this hotfix's validation.
After updating, **hard-refresh the browser (Ctrl+F5)** to load the JavaScript. Restart ComfyUI normally to update the displayed package version. A 3.9.1 Run/Take does not require migration for this UI hotfix. Versioned downloads are listed in [GitHub Releases](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases); Registry publication is separate.
## 3.9.1 — what changed (2026-10-04)
This patch fixes Second Pass conditioning, Take reuse and Review audio:
- **Second Pass:** preserve audio-only/mixed keyframes and inherit the verified conditioning and Reference assignments of the actual Review output, including a single middle chunk.
- **Take reuse:** include sampler closures and MODEL CFG/wrappers in compatibility checks to prevent reuse under different generation settings.
- **Review Driving Audio:** cut the source PCM once at the physical group's natural time, with the same start position for Exact ON/OFF.
- **Audio resampling:** prefer ComfyUI Core's standard API, with the legacy fallback for older Core versions. The Reference Encode Cache limitation for direct VAE weight changes is documented below.
**Updating from 3.9.0:** restart ComfyUI and refresh the browser after updating. Existing v5 Takes and history remain readable and are not deleted, but cannot be reused as prefixes for new v6 generation. Start a new Full Video or choose **Start again from Chunk 1**. Public node IDs, widget order and official workflows remain unchanged; keep using the **V3.9** nodes.
The repaired code passed **1,696 CPU tests / 1 skipped**, targeted Review Second Pass GPU checks and an integrated generation/reuse check. Browser acceptance remains pending. Optional latent-upscale doubled outlines are tracked separately in [Issue #27](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/issues/27); they are not a required gate for these fixes. This is a main-source update, not a new GitHub Release or Registry publication.
## 3.9.1 repair details
Second Pass preserves audio-only keyframes and adapts only the video component of mixed keyframes. Review Driving Audio uses the retained natural frame range of the physical decode group, measured before projection. Exact ON/OFF share that PCM origin. Completely empty Review intervals receive silence for the selected video duration; partially short audio stays short with Exact OFF.
New generation uses Sampling Contract v6 and Graph Contract v4, including sampler code/closures and MODEL CFG/wrapper settings. Saved v5 Takes, raw tensors and provenance remain readable, but cannot be reused in v6 generation. Start a new Full Video or use Review's **Start again from Chunk 1** to create a new revision. Old Takes remain intact. Unobservable external settings still allow fresh generation with reuse disabled; incompatible explicit Take/Retry requests are not replaced with a different Take.
Reference/Driving resampling prefers Core's standard API and uses legacy TorchAudio only when that API is absent. Preserved output PCM keeps its original waveform and sample rate.
**Reference Encode Cache limitation:** direct weight mutation inside the same VAE object is not detected automatically. Call `v3.ref_encode_cache.clear_ref_encode_cache()` afterwards or reload into a new VAE object. Unchanged VAE/input pairs continue to hit the cache. Decode Cache Helper reset affects a separate cache and does not clear Reference Encode Cache.
This repair stays within V3.9: public node IDs, socket/widget order, official workflows, Run Storage v3 and State/Session formats are retained. LoRA Plan is not included. CPU checks and browser/GPU acceptance are separate.
V3.9 also projects captured conditioning onto the actual Review output groups before Second Pass. A partial prefix or a single middle chunk can inherit its own conditioning and Reference assignments when the context validates. Full-sequence completion and saved Takes are unchanged; missing or mismatched context keeps the diagnostic prompt-only fallback.
The repaired runtime passed **1696 CPU tests, 1 skipped**, and three targeted GPU cases at **512×512 / 24 fps**: a two-group partial prefix, Chunk 2 alone, and complete three-group reuse. All groups inherited verified conditioning; saved FLAC matched the corresponding source PCM exactly and existing Takes/raw files were preserved. This validates the targeted repair. Browser acceptance remains pending; optional latent-upscale double outlines are a separate investigation in Issue #27. Historical V3.8 releases and tags remain available unchanged.
![H3 Continuum V3.9: fixed Reference Image slots, per-chunk assignments, and one Sampler input](images/v39/v39-feature-summary.png)
*V3.9 is current on `main`: fixed Image 1–9 slots, per-chunk Reference selection, and one Reference Images connection to the V3.9 Sampler.*
## Getting started video
[H3 Continuum V3.9 — Getting Started: Chunks, Prompts & Reference Images](https://youtu.be/AKJxBXaiG6Q) (12:26)
Learn basic First Image setup, chunk settings, Review/Take retries, Timeline prompting, and Reference Images. Includes English narration and subtitles, seven timestamped chapters, version-labeled UI stills, and actual generated examples.
![H3 Continuum V3.8X2 feature card from its original release](images/v39/v38x2-feature-summary-historical.png)
*This V3.8X2 card is retained for comparison. Its “CURRENT MAIN” label describes the time it was made; V3.9 is current now. The V3.8 Sampler and V3.8X2 workflow remain available, but their Reference wiring is not automatically converted to V3.9.*
## External INT Width/Height fix on main (2026-10-01)
Fixed the visible Width/Height input binding in both V3.8 and V3.9. External INT connections now stay aligned with their displayed controls through workflow save/reload and tab switching. Connected controls are read-only; their displayed numbers are saved Manual reserve values, not a prediction of the upstream node's runtime output. Disconnecting restores local editing.
To set exact dimensions from Core **Int** nodes, choose **Size Source = Manual** and connect the Width and Height INT inputs. **First Image** still uses the image and Resolution setting; connected dimensions do not override it or automatically change Size Source. If no active First Image is available, the existing Manual fallback applies. Dimension validation and backend input order/defaults are unchanged.
Validation: **1569 CPU tests passed, 1 skipped**, including 22/22 new frontend regression cases and 60/60 existing Review cases; Chrome save/reload and five tab round trips for both versions. One V3.9 GPU smoke run with external INT values produced exactly **480×640, 96 frames at 24 fps, with 4-second audio** in **107.3 s** (20 sampling steps); the entire saved video decoded successfully. This is a short single-condition check, not long-form or all-model acceptance. Sampling, Run Storage, the earlier Review/cache repairs, official workflow JSONs/ZIPs and historical Releases are unchanged.
Update Continuum from `main`, then **hard-refresh the browser (Ctrl+F5)** to load the new JavaScript. This fix alone does not require a ComfyUI backend restart; restart normally if the update also brings Python changes or new node definitions. This remains a main-source fix, not a new Release or Registry package.
## V3.9 performance fix on main (2026-10-01)
V3.9 could repeat unnecessary prompt-conditioning preparation when **Prompt Format = Fixed** and **no Reference Images are selected**. This repair reuses the existing bounded conditioning cache when its inputs are unchanged. First Image is supported; List/Timeline prompts, selected Reference Images, Reference Audio, Timeline Video and terminal-pair groups keep their existing paths. This is not the Decode Cache Helper and does not skip Sampling.
In the tested `2 × 5 s`, 480×640 First Image workflow, group preparation fell from about **10.76 s to 0.013 s**. In the post-fix same-session comparison (three measured runs per version, after warm-up), median total time was **273.58 s for V3.8 vs 273.78 s for V3.9**; the ranges overlapped. The earlier roughly 6% gap was no longer observed under these conditions. This does **not** establish a speed fix for every workflow, model or GPU, or a GPU-kernel speedup.
Update Continuum from `main` and **fully restart ComfyUI** to load the Python changes. Validation: **1558 CPU tests passed, 1 skipped; all 8 GPU generation runs succeeded** on ComfyUI 0.38.0 / RTX 5060 Ti 16 GB. Seeds, chunk/frame contracts, Run Storage, V3.8X2/V3.9 official workflows and ZIPs, and historical Releases are unchanged. This is a main-branch fix, not a new Release or Registry package.
## Review restoration fix on main (2026-10-01)
Switching to another workflow and back could hide Review actions even though saved Takes still existed. This frontend repair waits for workflow restoration before reading history and ignores stale results from previous graphs. Real generation-setting changes still disable Review/Take application; restoring the original settings re-enables the actions.
Update the custom node from `main`, then refresh the browser to load the new JavaScript. Validation: 60/60 frontend regression cases, 108 focused CPU tests, five Chrome tab round trips, and a Chrome Base Seed edit/restore test; saved Take metadata stayed unchanged. Sampling, Run Storage contracts, V3.8X2/V3.9 workflow files and historical Releases are unchanged. GPU continuation was not tested for this repair. [Details and verification limits](REVIEW_RESTORATION_REPAIR.md).
## Archived implementation notes (pre-V3.8; not the current guide)
The version-labelled sections below are retained temporarily as source history only. They do not describe the V3.8 standard workflow, public surface, or recommended wiring. For an older workflow, use the migration policy and its matching historical Release/tag.
### V3.7 High-Resolution Refinement Foundation
V3.7 adds **H3 Continuum Sampler V3.7** and completes two foundations for resolution-changing Second Pass workflows without changing V3.6 Production defaults or the existing SIGMAS socket.
### Conditioning survives a resolution change
The Conditioning Adapter rebuilds First/Last images, continuation context, and an owned Still Image Guide at the target latent geometry. Physical-group ownership, Long Terminal Merge, Core PackedLayout/RoPE reconstruction, and first-pass Audio passthrough are preserved. Correctness gates passed at 704×704 and 1152×1152; larger canvases still require substantially more memory and processing time.
### RefineSchedule gives refinement ranges explicit meaning
The internal, versioned `RefineSchedule` contract supports `External`, `Full`, `Tail`, and `Partial`. `External` preserves the incoming SIGMAS tensor exactly, while `Tail` and `Partial` select exact ranges from the same source schedule. The accepted internal Tail 6 gate reproduced the six-evaluation suffix and preserved decoded Audio PCM bit-exact. This contract is an implementation foundation rather than a new public widget, and Production defaults remain unchanged.
### Tail 6 cuts Second Pass time by about one third
The largest practical benefit in V3.7 is a shorter high-resolution Second Pass when the internal schedule uses Tail 6. Instead of evaluating all ten points, Tail 6 runs the exact low-sigma suffix that performs the final refinement work.
| Second Pass range | Sampling time |
|---|---:|
| Tail 10 | **100.27 s** |
| Tail 6 | **64.18 s** |
| Saved | **36.09 s / 36.0%** |
In this gate, Tail 6 finished in about 64% of the Tail 10 time: roughly 1 minute 4 seconds instead of 1 minute 40 seconds. A separate Learned Latent Upscaler gate measured `121.17 s -> 66.32 s` for Tail 10 versus Tail 6, about 45% lower, while API total changed from `170.70 s` to `119.61 s`, about 30% lower. These measurements are configuration-specific and are not universal speed guarantees.
`RefineSchedule` does not make each model evaluation faster. It gives Continuum an exact, versioned way to manage and execute the Tail 6 range; the speedup comes from running six evaluations instead of ten. Tail 6 is not enabled automatically, and V3.7 does not change Production defaults.
### Still Image Guide remains Experimental
V3.7 can map one Still Image Guide from an absolute frame to its owning physical group and rebuild it from the original image during a high-resolution Second Pass. Its positioning, Terminal Merge ownership, Run Storage identity, and non-owner isolation passed correctness testing. Core Add Guide uses hard-anchor semantics, however, so a Guide can cause an abrupt trajectory change at the anchor and can redirect later motion. **Still Image Guide is Experimental and remains on Production HOLD.** It should not be treated as a smooth transition control.
### V3.6.1 Maintenance Hotfix
V3.6.1 fixes focused fail-open/compatibility issues without changing the accepted Balanced 22 Masked AV path. Mixed Timeline/List prompts now report `H3C-P105`; Timeline parsing ignores only standalone `---` lines, keeps existing uncovered-chunk fallback, and never stops generation because of prompt syntax.
With `Continuation Backend = Standard` and Audio Continuity enabled, `Balanced — 22 frames` continues to use Masked AV. `Fast — 5 frames`, `Strong — 39 frames`, and `Auto — conservative` now resolve to the accepted Reference Context transport before Run Storage identity is created. The selected UI values are not rewritten, and the status report states the resolved transport and reason. Standard with Audio Continuity disabled remains Masked Video; Compatibility remains Reference Context.
Run Storage now treats every absent Last Frame spelling as the same empty identity. A normal no-Last-Frame run can therefore extend from two to three chunks as `2 reused, 1 generated`, including narrowly compatible V3.6.1 caches whose former final chunk stored `none`. A real connected Last Frame and the Long Terminal Merge atomic pair remain strict. Turning Run Storage Off also resets the hidden Regenerate From and Variation Nonce values to `Auto` and `0` before queueing.
### V3.6 Masked AV Continuation
![H3 Continuum V3.6 Masked AV Continuation](images/v36-masked-av-continuation.png)
V3.6 adds **H3 Continuum Sampler V3.6**. Its Standard continuation backend places the previous finalized Video and Audio latent prefixes directly inside the next H3 target and protects them with Core noise masks. Unlike the V3.5 Reference Context route, it does not append the same prior frames as a separate Reference block, reducing the packed sequence processed by H3.
| Continuation Backend | Purpose |
|---|---|
| `Standard` | Recommended V3.6 path. Balanced 22 + Audio Continuity uses Masked AV; Fast 5, Strong 39, and Auto safely fall back to Reference Context. When Audio Continuity is off, Masked Video is used. |
| `Compatibility` | Advanced fallback using the accepted V3.5 Reference Context behavior. |
Internal transport identifiers are not exposed in the UI. Run Storage identity follows the resolved transport: Balanced 22 Masked AV stays separate from Reference Context, while a non-Balanced Standard fallback safely shares the identical accepted Reference contract. `Regenerate From` treats the Long Terminal Merge logical pair `[2,3]` as one atomic physical sample.
### Accepted behavior
- T2VA, I2VA, Reference Image + Reference Audio, Balanced 22-frame Joint AV, and 3×5-second FL2VA Long Terminal Merge GPU gates passed.
- Protected Video and Audio prefixes finalize bit-exact; generated regions remain owned by the sampler.
- The accepted Reference Image + Reference Audio output passed subjective audio listening with no reported click, dropout, or stereo-positioning defect.
- In the tested FL2VA Terminal Group 2 pairs, Standard removed 2,874 packed rows (`8.01%`) and reduced median Sampling time by `4.06%` versus Compatibility. Performance varies by model, hardware, and workflow.
- `chunk_seconds` now accepts 4.0–30.0 seconds with a 5.0-second default and 0.1-second step. The 5–15 second range remains recommended and validated; longer high-resolution chunks can substantially increase VRAM use and runtime.
In the historical V3.6 package, V3.5.3, V3.5, and V3.4 Node IDs remained registered for saved workflows. V3.5.3 retained its original Reference Context behavior rather than being silently redirected to the V3.6 backend. This is not a V3.8 registration guarantee.
## V3.5.3 Maintenance Hotfix
V3.5.3 is a distribution-integrity maintenance release. It repairs the public Conditioning Bridge workflow's Width/Height links, removes stale orphan link IDs from the published V3.4/V3.5 templates, corrects Hybrid `<Picture N>` warning numbering, and repairs one legacy PNG for strict image decoders.
There are no changes to generation behavior, nodes, sockets, Sampling, Conditioning payloads, Terminal Merge, Assembly, Seam, Run Storage, Prompt/CLIP caching, Video Guide optimization, or V3.4 compatibility. Existing V3.5.x workflows continue to load unchanged.
## V3.5.2 Stabilization & Optimization Update
V3.5.2 adds no new generation mode. It stabilizes the V3.5.x release, removes repeated Prompt/CLIP work when safe, and lowers temporary memory for long Video Guide inputs while preserving existing workflow and generation contracts.
![H3 Continuum V3.5.2 stabilization and optimization results](images/v352-stabilization-optimization.png)
The figure records the accepted optimization measurements. The final packaging gate additionally passed on ComfyUI 0.33.3 with a 172-entry Manifest, including this image.
### What changed
1. **Repeated-run Prompt/CLIP cache** — unchanged T2VA, I2VA, and FL2VA conditioning can reuse CPU-cached Prompt/CLIP results. Reference, scheduled, tokenizer-option, and hook-modified CLIP paths conservatively bypass the cache. Sampling still runs normally.
2. **Video Guide memory optimization** — the complete source remains part of finite validation and SHA-256 identity, but only the required prefix remains stored for VAE/conditioning work.
3. **Stabilization audit** — measured Sampling, Decode, Assembly, saving, optional-feature tax, and retained memory. No speculative Sampling, Driving Audio, Audio hash, Assembly, Seam, Session, or V3.4 optimization was adopted.
### Measured results
| Area | Before | V3.5.2 | Result |
|---|---:|---:|---:|
| T2VA repeated Prompt/CLIP | 5.898398 s | 0.000028 s | Cache HIT, `encode_calls=0` |
| FL2VA repeated Prompt/CLIP | 21.642687 s | 0.007128 s | Cache HIT, `encode_calls=0` |
| Video Guide peak additional RSS | 396.8 MiB | 114.7 MiB | About 71% lower |
| Video Guide retained storage | 225 MiB | 93 MiB | About 59% lower |
Prompt/CLIP figures measure only the conditioning subphase, not total generation time. The Video Guide memory benchmark uses deterministic 300×256×256 RGB input with a 124-frame required prefix. A separate GPU A/B used an actual 15-second / 360-frame Video Guide and produced identical decoded video and audio PCM.
The measured Sage-only production baselines on the tested RTX 5060 Ti 16 GB / 64 GB system were 168.069 seconds for 1×5-second 576×576 T2VA and 379.765 seconds for 3×5-second 640×640 FL2VA Long Terminal Merge. These are configuration-specific baselines, not universal speed guarantees. Sampling remained the dominant cost; Continuum Assemble + Seam stayed below 1%.
> **V3.8.0 is the historical release baseline.** V3.8X2 remains a separate package-3.8.3 workflow, while V3.9 source is available on `main`. Historical implementation modules remain in source because V3.8X2 reuses them internally. The current registration has ten V3.8X2 IDs plus two V3.9 IDs. Use the matching historical Release/tag for workflows requiring other IDs. Still Image Guide remains Experimental.
## V3.5.1 Reference Audio & Compatibility Update
V3.5.1 added two focused features without changing the V3.4 Sampling, Conditioning, Terminal Merge, Assembly, Seam, or Run Storage contracts:
1. **Optional Reference Audio** — native H3 audio conditioning that does not replace the generated final audio.
2. **Conditioning Bridge V3.5** — one complete Core-compatible `MODEL` and `CONDITIONING` object per Continuum physical group for external sampler workflows.
The Reference Audio sockets are permanently defined by Python `INPUT_TYPES`. Dynamic socket changes and the UI-only `Hidden / Show` control were removed to prevent workflow save/reload value shifts. Node IDs, backend keys, Sampling, Conditioning, standard Seed handling, and other widgets are unchanged.
In the historical V3.5.1 package, V3.4 node IDs and backend socket keys remained registered for saved-workflow compatibility. V3.4/V3.5 workflows continued to load there; V3.5.1 clarified the displayed input names without rewriting saved links. V3.8 instead uses the seven-node support boundary above.
### Reference inputs at a glance
The displayed names intentionally describe different jobs. They are not interchangeable pairs:
| Displayed input | Connect from | Purpose | Final audio behavior |
|---|---|---|---|
| `reference_image_1`–`reference_image_3` | Still-image `IMAGE` | Persistent identity, subject, or appearance references | No audio |
| `Video Guide Frames` | Video loader `IMAGE` frame batch | Persistent motion, framing, timing, and appearance guidance across chunks | Does not carry the video's audio |
| `Driving Audio` + `Driving Audio VAE` | Audio loader, or video loader `AUDIO`, plus the matching audio VAE | Audio guidance whose effective source stream is preserved for final output | Replaces generated final audio with the selected source stream |
| `Reference Audio (Optional)` + `Reference Audio VAE (Optional)` | Standalone audio plus the matching audio VAE | Native H3 conditioning only | Does not copy or replace the generated final audio |
For the common video-with-sound case, connect the loader's `IMAGE` output to `Video Guide Frames` and its `AUDIO` output to `Driving Audio`. Do not connect that audio to `Reference Audio (Optional)` unless conditioning-only behavior is specifically intended.
![Sampler V3.5 permanent Reference Audio, Video Guide Frames, Driving Audio, and Reference Image inputs](images/v351-video-guide-frames.png)
`Reference Audio (Optional)` and `Reference Audio VAE (Optional)` are always present exactly as defined by the node's Python input schema. V3.5.1 no longer adds or removes these sockets dynamically and has no frontend-only `Hidden`/`Show` widget. This keeps positional widget values aligned when workflows are saved and reloaded; backend input keys and existing workflow links are unchanged.
![Sampler V3.5.1 with permanent Reference Audio sockets and aligned widgets](images/v351-reference-audio-permanent.png)
### Conditioning Bridge V3.5
`H3 Continuum Conditioning Bridge V3.5` is the Advanced connection point for external sampling. Connect `model`, `clip`, the externally processed `video_latents`, `assembly_plan`, and `refine_context`; connect `video_vae` only when the selected conditioning path requires it. The node returns parallel `group_models` and `conditioning` lists whose length equals the physical-group count. Each list item remains a complete ComfyUI object—conditioning entries are never flattened into the physical-group list.
```text
H3 Continuum Sampler V3.5.video_latents
-> external H3 latent processor (for example LBH)
-> H3 Continuum Conditioning Bridge V3.5
-> Core BasicGuider + external sampler
-> Core Video / Audio VAE Decode
-> H3 Continuum Assemble + Seam V3.5
```
The external workflow owns AV LATENT pairing, noise, SIGMAS, Audio Lock, sampling, and audio passthrough. The Bridge only exposes the prepared `MODEL` and `CONDITIONING` objects plus the updated Assembly Plan.
![V3.5.1 LBH and Conditioning Bridge external sampling flow](images/v351-lbh-conditioning-bridge-flow.svg)
The retired V3.5.1 LBH + Conditioning Bridge example is available only from its matching historical Release/tag; it is not a V3.8 workflow.
The example keeps the standard Core titles `BasicGuider`, `BasicScheduler`, and `SamplerCustomAdvanced`. Public examples do not rename Core nodes, so they remain immediately distinguishable from Continuum and third-party nodes. The workflow also uses external nodes for LBH latent upscaling, AV LATENT concatenation/separation, loading/saving, and optional acceleration; install or replace those nodes according to your ComfyUI environment.
LBH changes latent geometry only. It has no Continuum denoise-strength control: use the external `BasicScheduler` SIGMAS to decide how strongly the resized latent is regenerated. A larger LBH scale such as 1.5x may still be fast, but memory, decode, and sampling costs increase with the target canvas.
## What's new in V3.5
V3.5 has two major additions:
1. **Continuum-aware Second Pass / Hi-Res Fix** — refine externally processed H3 latents or use the integrated pixel/VAE 2x path without changing V3.4 sampling.
2. **Low-memory Assemble + Seam** — write the final video IMAGE directly to RAM or a Windows-safe mapped file while preserving Exact Duration, Seam, Terminal Merge order, and audio behavior.
### 1. Second Pass and Hi-Res Fix
**H3 Continuum Hi-Res Fix V3.5** is the one-node Main path. It performs Video VAE Decode, bounded pixel resize, Video VAE Encode, and one low-denoise Second Pass. The integrated Main path remains **Experimental** because long 2x runs can exceed GPU memory.
![H3 Continuum Hi-Res Fix V3.5 node](images/v35-hires-fix-node.png)
Normal Hi-Res Fix wiring:
```text
H3 Continuum Sampler V3.5
-> H3 Continuum Hi-Res Fix V3.5
-> Core Video / Audio VAE Decode
-> H3 Continuum Assemble + Seam V3.5
```
**H3 Continuum Second Pass V3.5** is the stable Advanced bridge requested in Issue #8. It accepts an externally processed H3 video latent and samples each Continuum physical group once while preserving prompt policy, seed reporting, group order, and the original first-pass audio LATENT object.
![H3 Continuum Second Pass V3.5 node](images/v35-second-pass-node.png)
External latent processor wiring, such as LBH:
```text
H3 Continuum Sampler V3.5.video_latents
-> external H3 latent processor
-> H3 Continuum Second Pass V3.5
-> Core Video / Audio VAE Decode
-> H3 Continuum Assemble + Seam V3.5
```
Connect the original `audio_latents`, `assembly_plan`, and `refine_context` from Sampler V3.5 to the Hi-Res Fix or Second Pass node. Use the same final MODEL and CLIP, plus a separate sampler and workflow-side SIGMAS for the refinement pass. The nodes do not contain an internal denoise value.
The `BasicScheduler` is deliberately omitted from the screenshots. It is still required to create the Second Pass `SIGMAS`. The GPU-tested baseline is:
```text
sampler: res_multistep
scheduler: simple
steps: 10
denoise: 0.35
```
Do not chain Hi-Res Fix and another Second Pass in the standard path: Hi-Res Fix already performs one Second Pass internally. `H3 Continuum Latent Resize V3.5` is an advanced utility, not the recommended Main 2x path; direct high-ratio interpolation of the 24-channel H3 latent produced persistent artifacts in testing.
If Hi-Res Fix is not connected, V3.4 sampling remains on its original path. V3.5 does not add a forced CPU-offload or low-VRAM sampler mode. `Hi-Res Fix enabled=false` is a lazy exact passthrough and does not evaluate its resize, VAE, or Second Pass dependencies.
See [V3.5 Hi-Res Fix and Second Pass workflow guide](V35_HIRES_FIX.md) for exact sockets, bypass behavior, external-upscaler wiring, and limits.
### 2. Low-memory Assemble + Seam
**H3 Continuum Assemble + Seam V3.5** adds manual RAM / Disk-backed selection and a conservative Auto policy. Disk-backed mode writes the final video IMAGE directly to a mapped file instead of rebuilding the entire completed video in anonymous RAM. Audio remains in RAM.
| Backend | Behavior |
|---|---|
| Auto | Uses RAM only when the final IMAGE is at most 4 GiB and physical-memory reserve is sufficient; otherwise selects Disk-backed. Missing memory counters fail safely to Disk-backed. |
| RAM | Existing in-memory output behavior. |
| Disk-backed | Maps the final video IMAGE to a temporary file; video only. Audio remains in RAM. |
Disk-backed mode preserves the backing file while ComfyUI still references the returned Tensor. Downstream nodes can still allocate a full RAM copy, so Continuum's low-memory guarantee applies to its own assembly stage, not arbitrary downstream implementations.
#### Measured memory reduction
The validated stress case assembled a deterministic `1536 x 1536`, 360-frame float32 IMAGE with a final size of `9.49 GiB`. RAM and Disk-backed outputs were video/audio hash-identical.
| Windows process metric | RAM backend | Disk-backed | Reduction |
|---|---:|---:|---:|
| Private memory | 12.41 GiB | 2.91 GiB | **9.50 GiB** |
| USS | 10.29 GiB | 0.80 GiB | **9.49 GiB** |
This is a **system-RAM/private-commitment** reduction during Continuum Assembly, not a claim that sampler GPU VRAM is reduced. Windows RSS includes resident mapped-file pages and was therefore similar between the two backends; private memory and USS are the relevant measurements. Disk-backed I/O can make final assembly slower, but it does not slow the preceding model sampling path.
#### Optional Windows tip: disable pinned memory
For MiniMax H3 on Windows, starting ComfyUI with `--disable-pinned-memory` can substantially reduce host-RAM pressure. In one local RTX 5060 Ti 16 GB / RAM 64 GB check, a `0.4 MP (640 x 640), 2 x 5s` run completed in `5m 23s`, with about `6.1 GB` sampler RSS and `8.6 GB` after assembly, while showing little practical speed difference from normal startup.
```text
--disable-pinned-memory
```
This is an environment-specific tip, not a default requirement or a universal crash fix. Compare the same workflow and seed with only this option changed, and keep it only when memory use improves without a meaningful speed or stability regression. It disables ComfyUI's pinned host-memory path; it does not guarantee that Windows Shared GPU Memory will never be used.
### GPU acceptance at a glance
| Path | Tested condition | Result |
|---|---|---|
| Main Hi-Res Fix | FL2VA 1 x 5s, 576 -> 1152 | PASS |
| Main Hi-Res Fix | Hybrid FL2VA + Reference 1 x 5s, 576 -> 1152 | PASS |
| Advanced Second Pass | Hybrid/Reference 1 x 5s, 576 x 576 | PASS |
| Main Hi-Res Fix | FL2VA 3 x 5s, 576 -> 1152, RTX 5060 Ti 16 GiB | **CUDA OOM at terminal 77T group** |
For the failed 3 x 5s case, First Pass and the 37T Second Pass group completed. The final logical chunks `[2,3]` remained one 77T Long Terminal Merge physical group and exceeded the tested 16 GiB GPU at its first 1152x1152 Second Pass inference. This is a documented resource limit; the Continuum grouping contract did not change.
## V3.4 compatibility baseline
The historical V3.5 package retained V3.4 nodes for saved workflows; it did not replace their Node IDs, public sockets, Sampling, Conditioning, Terminal Merge, Assembly, Seam, or Run Storage behavior. Those historical workflows remain available through the corresponding Release/tag. Current V3.9 exports only its twelve registered IDs, even when older implementation modules remain in source.
![H3 Continuum V3.4 workflow overview](images/v34-workflow-overview.png)
V3.4 focuses on the two reference workflows that are most useful in normal production:
- **Driving Audio**: use an existing audio source as the preserved final audio while it guides the H3 generation across chunks.
- **Video Guide Frames**: use a video loader's IMAGE frame batch as persistent guidance for appearance, motion, framing, and timing without treating it as a frame-by-frame copy.
- **Hybrid FLF + Reference**: use First Frame, Last Frame, and persistent Reference Images together without adding a separate mode or model allowlist.
- **FL2VA Terminal Merge**: keep the final two 5-second FL2VA chunks as one physical 10-second sampling and decode unit for Core-equivalent Last Frame handling.
- **Restartable chunks**: reuse completed chunks with Run Storage and regenerate only the selected part when the generation contract remains compatible.
- **Core compatibility**: remove Continuum-only rejection of unknown upstream/custom nodes and remove the obsolete `strict_compatibility` control.
- **Spectrum interoperability**: use the official H3 Continuum Interop API v1 when Spectrum is installed; Spectrum remains optional.
- **Simpler public interface**: Timeline Video and the earlier timeline-audio paths are hidden from the V3.4 public sampler interface rather than being presented as stable production features.
### Direction change: from timeline generation to preserved references
Earlier development explored Timeline Video and timeline-audio conditioning. Those paths can be useful for experiments, but they are not the default production behavior: chunk boundaries can change visual content, and generated audio can diverge from a supplied song, dialogue, or effects track.
V3.4 therefore prioritizes:
1. **Driving Audio** for cases where the supplied audio should remain the final audio stream.
2. **Video Guide Frames** for cases where a supplied video should guide the generated result without requiring exact frame reproduction.
3. **Chunked generation and Run Storage** for practical retries and long-form work.
This is a usability and reliability decision, not a claim that the experimental timeline paths are impossible. They are hidden from the stable interface while the public workflow stays focused on predictable user-controlled inputs.
## Example workflow
- [V3.9 workflow JSON](../examples/workflows/MiniMax_H3_Continuum_V39.json) and [ZIP](../examples/workflows/MiniMax_H3_Continuum_V39.zip) — current V3.9 graph; the ZIP contains only this JSON
- [V3.8X2 workflow JSON](../examples/workflows/MiniMax_H3_Continuum_V38X2.json) — retained V3.8X2 graph; Spectrum is disabled by default
- [V3.8X2 workflow ZIP](../examples/workflows/MiniMax_H3_Continuum_V38X2.zip) — contains both the V3.8X2 and V3.9 JSON files; it is not a custom-node installer
- [V3.8X2 Helper-explicit JSON](../examples/workflows/MiniMax_H3_Continuum_V38X2+Decode_Cache_Helper.json) and [ZIP](../examples/workflows/MiniMax_H3_Continuum_V38X2+Decode_Cache_Helper.zip) — the same graph under an explicit distribution name
The declared Registry payload includes both V3.9 and V3.8X2 graphs; it is **not dependency-free**. Spectrum, rgthree, and KJNodes must be installed separately; ComfyUI-Easy-Use is not required. Historical workflow files remain in GitHub source but are excluded from the Registry payload. For an older saved workflow, use its matching historical Release/tag as described in the [migration policy](V38_RELEASE_AND_MIGRATION.md). Registry packaging validation/publication is separate from this GitHub source update.
## V3.4 feature details
### Driving Audio
Driving Audio is the primary V3.4 audio path.
- Source audio guides each chunk at its absolute sequence position.
- The original effective stream is preserved for final output.
- Generated audio and Audio Seam processing are bypassed while Driving Audio is active.
- Audio is not independently rewritten at every chunk boundary.
- Perfect lip sync is not guaranteed; results still depend on H3, source material, prompts, and sampling.
### Video Guide Frames
`Video Guide Frames` provides a persistent native H3 video guide across all chunks. Despite its ComfyUI `IMAGE` socket color, it expects the frame batch produced by a video loader rather than one ordinary still-image reference.
- It guides motion, framing, timing, and appearance.
- It is not a direct pixel-copy or deterministic identity-transfer system.
- Video and audio are independent. Route a loader's IMAGE output to `Video Guide Frames` and AUDIO output to `Driving Audio`.
- Source decoding and frame-rate conversion remain loader responsibilities. Continuum does not impose an unnecessary forced 24 fps conversion.
| Video Guide Size | Purpose |
|---|---|
| Efficient - 0.4 MP | Default; lower token cost and faster iteration. |
| Balanced - 0.6 MP | More detail at a higher compute cost. |
| Match Output | Largest reference input; potentially much slower and heavier. |
#### Conditional size controls and internal resize
The compact V3.5 UI shows each size control only when its corresponding input is connected:
- Connect any of `reference_image_1` through `reference_image_3` to reveal `Reference Size`.
- Connect `Video Guide Frames` to reveal `Video Guide Size`. Earlier workflows and reports may call this control `Video Reference Size`; V3.5.1 uses the clearer display name `Video Guide Size` while preserving the existing backend key and saved-workflow compatibility.
These controls are therefore not missing from a newly created node; they are hidden until relevant. Reference Images and Video Guide Frames are resized internally before VAE encoding. Video Guide Frames preserve their aspect ratio, align to the H3 canvas, and use Lanczos when reduction is needed. Smaller sources are not enlarged automatically. An external resize node is normally unnecessary, but remains useful for deliberate cropping, forced upscaling of a small source, or a custom resize/upscale algorithm. Video decoding and frame-rate conversion remain the loader's responsibility.
### Hybrid First/Last Frame + Reference Images
You need an hybrid version of MiniMax-H3 to make both I2V and reference images work:
https://huggingface.co/smhfacct/Minimax-H3-fl2va-ref2va-hybrid-models
https://www.reddit.com/r/StableDiffusion/comments/1vm62pj/i_tested_out_the_b2049_hybrid_variant_for_r2v/
V3.4 supports First Frame, Last Frame, and persistent Reference Images in the same Continuum run. The public sockets and controls are unchanged: connect the inputs that the selected H3 model supports.
- Pure FL2VA and pure Ref2VA identity contracts remain unchanged.
- The combined path receives a distinct hybrid identity so Run Storage does not reuse incompatible chunks.
- Prompt/reference mismatches produce guidance warnings rather than a Continuum-only execution stop.
- No model-name allowlist or automatic model replacement is used.
Local checks passed with the B2049 hybrid variant. FL-only models could run, but showed less reliable transitions when FLF and Reference Images were combined; a hybrid-capable model is therefore recommended for this path.
### FL2VA Terminal Merge
For 5-second FL2VA workflows, Continuum keeps the final two logical chunks together as one physical sampling and VAE decode unit.
- `2 x 5s`: one 243-frame physical sample, decoded once, then adjusted to the requested 240-frame output.
- `3 x 5s` and longer: earlier chunks use normal Continuation; the final two chunks use one 260-frame physical sample containing the 22-frame continuation context.
- After the terminal physical decode, the 22-frame context prefix is removed. For `3 x 5s`, this gives `124 + 238 = 362` frames before exact-duration adjustment to 360 frames.
- If the final two prompt sections differ, they are rebased internally to a local `[0-5s]` / `[5-10s]` timeline. Identical prompts retain the accepted shared-prompt path.
- The terminal pair uses one physical seed and one physical decode while Run Storage continues to preserve normal logical chunk entries.
Local GPU comparison confirmed that `2 x 5s` matches Core 10-second FL2VA sampling and terminal decode behavior. A `3 x 5s` Timeline run also passed with three logical chunks, two physical sampling passes, two physical decode groups, and a 15.000-second output. The short still period near the supplied Last Frame was also present in the equivalent Core result and is treated as normal FL2VA convergence rather than a Continuum-specific failure.
### Core-first, permissive execution
V3.4 removes or relaxes Continuum-only rejection rules that blocked otherwise runnable Core H3 experiments.
- Prompt problems prefer warnings and fallback behavior instead of stopping.
- Unknown upstream/custom wrapper classes are not blanket-rejected.
- There is no model allowlist and no automatic model replacement.
- The obsolete strict_compatibility control is removed from the V3.4 interface.
- Current Core PackedLayout behavior is supported without requiring the removed legacy frame_count argument.
Hard stops remain only for states that cannot execute safely, including corrupt latent topology, incompatible assembly data, invalid persisted revisions, or unusable mandatory payloads.
### Cleaner interface
The public nodes focus on normal production controls. Developer diagnostics and detailed reports are available through ComfyUI settings. The current release path uses `H3 Continuum Sampler V3.8`; older sampler and assembler notes below are retained for historical compatibility context.
<img width="389" alt="H3 Continuum Sampler V3.7 with optional Guide input" src="images/v37-sampler-node.png" />
![H3 Continuum Assemble + Seam V3.5](images/v35-assemble-seam-node.png)
### Timeline paths
Experimental Timeline Video and earlier timeline-audio paths remain in historical implementation code but are not exported by V3.8. Open V3.3 workflows with their historical package; new V3.8 workflows should use `Driving Audio` and `Video Guide Frames`.
## Installation
V3.8 is verified against ComfyUI 0.34.2. Historical acceptance records below may mention older ComfyUI versions; they are not the current installation target.
### Updating an existing installation
For a Git checkout, run the update from the custom-node directory:
```powershell
cd ComfyUI/custom_nodes/ComfyUI-H3-Continuum
git pull --ff-only origin main
```
Restart ComfyUI after the update. If the node was installed with ComfyUI Manager, use its **Update** action instead of running `git pull` manually. Do not mix Manager updates and a separate Git checkout for the same installation.
After the backend restart, search for `H3 Continuum Sampler V3.8` with the V3.8X2 workflow, or `H3 Continuum Sampler V3.9` and `H3 Continuum Reference Images V3.9` with the V3.9 workflow. The source registers twelve IDs in total. If they are missing, check the startup console for the `H3 Continuum 3.9.0 loaded` message and any `ComfyUI-H3-Continuum` import error.
Search for H3 Continuum or Continuum in ComfyUI Manager, or install manually:
~~~bash
cd ComfyUI/custom_nodes
git clone https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum.git
~~~
Restart ComfyUI after installation or update.
## V3.5 nodes
### H3 Continuum Sampler V3.5
The V3.5 sampler preserves V3.4 generation behavior and adds `refine_context` as the sixth output. Connect it to V3.5 Hi-Res Fix, Second Pass, or Conditioning Bridge paths that need the original conditioning context. V3.5.1 also provides permanently visible optional Reference Audio sockets.
### H3 Continuum Conditioning Bridge V3.5
Advanced external-sampler bridge. It returns aligned `group_models` and complete `conditioning` objects per physical group together with the updated Assembly Plan. It does not create AV LATENT pairs, noise, SIGMAS, Audio Lock, or external sampling policy.
### H3 Continuum Hi-Res Fix V3.5 (Experimental)
Main convenience node. When enabled, it performs the Video VAE pixel-resize round trip and Second Pass internally. When disabled, video/audio LATENT objects and the Assembly Plan pass through unchanged and the sampling/VAE dependencies remain lazy.
### H3 Continuum Second Pass V3.5
Advanced Issue #8 bridge for an externally processed video latent. It samples once per physical group, reads the physical prompt contract from the Assembly Plan, and returns the original first-pass audio LATENT objects.
### H3 Continuum Latent Resize V3.5
Utility that changes only video-latent H/W. It does not sample. Direct 2x interpolation at large square targets is not the accepted Main Hi-Res Fix method.
### H3 Continuum Assemble + Seam V3.5
V3.4-compatible assembly with Auto, RAM, and Disk-backed video-buffer backends. Exact Duration and seam corrections write directly to the selected target buffer.
## V3.4 nodes
### H3 Continuum Sampler V3.4
Main inputs:
- model, clip, video_vae, sampler, sigmas
- Sequence Prompt
- first_frame, last_frame
- reference_image_1 through reference_image_3
- Video Guide Frames
- driving_audio, audio_vae
Outputs:
- video_latents, audio_latents
- assembly_plan, status
- driving_audio
Visible controls:
- Prompt Format, chunks, chunk_seconds
- width, height, continuity
- base_seed, control after generate
- audio_continuity
- Run Storage
- Reference Size
- Video Guide Size
`chunk_seconds` uses one shared duration for every chunk. The default is 5.0
seconds; 5–15 seconds remains the recommended and validated range. Durations up
to 30.0 seconds are supported, but values above 15 seconds can substantially
increase VRAM use and processing time, especially at high resolution. Per-chunk
variable durations are not supported.
### H3 Continuum Assemble + Seam V3.4
Inputs: images, audio, assembly_plan, and driving_audio.
Controls: Audio Seam and Video Seam.
When Driving Audio is connected, preserved source audio is selected for final output and generated audio seam processing is bypassed.
## Connection order
~~~text
H3 model / CLIP / Video VAE / sampler / sigmas
|
v
H3 Continuum Sampler V3.4
| | |
video_latents | assembly_plan
|
audio_latents
video_latents -> Core VAE Decode ------- images --+
audio_latents -> Core VAE Decode Audio -- audio --+--> H3 Continuum Assemble + Seam V3.4
assembly_plan -------------------------------------+
driving_audio -------------------------------------+
~~~
For a source video with sound:
~~~text
Video loader IMAGE -> Video Guide Frames
Video loader AUDIO -> driving_audio
~~~
The supplied templates use Video Helper Suite for this split. Any compatible IMAGE/AUDIO loader may be substituted.
## Prompt formats
### Timeline
For **Timeline** mode, each time-range header must be written on its own line. Write the scene description on the following line or lines.
Recommended format:
~~~text
[0-5s]
First section: action, camera movement, and scene progression.
[5-10s]
Continue from the exact final state of the previous section.
[10-15s]
Continue naturally without resetting the scene.
~~~
For 10-second chunks:
~~~text
[0-10s]
Describe the first scene.
[10-20s]
Describe the next scene.
[20-30s]
Continue the sequence.
[30-40s]
Describe the final section.
~~~
**Important:** do not place the prompt text on the same line as the time header.
Correct:
~~~text
[0-5s]
Describe the scene.
~~~
Not recommended:
~~~text
[0-5s] Describe the scene.
~~~
The second form is not recognized as a Timeline header. When **Prompt Format = Auto**, it may therefore be interpreted as a Fixed prompt instead, causing the complete text to be reused across chunks.
The time ranges should normally match the configured `chunk_seconds`.
- `chunk_seconds = 5` → `[0-5s]`, `[5-10s]`, `[10-15s]` ...
- `chunk_seconds = 10` → `[0-10s]`, `[10-20s]`, `[20-30s]` ...
- `chunk_seconds = 15` → `[0-15s]`, `[15-30s]`, `[30-45s]` ...
- `chunk_seconds = 30` → `[0-30s]`, `[30-60s]`, `[60-90s]` ...
#### Global preamble and per-chunk prompting
Text placed before the first Timeline header is treated as a **global preamble** and is automatically included in every chunk prompt. This is a good place for information that should apply throughout the sequence, such as subject definitions, overall style, persistent environment details, or other global instructions.
However, the Timeline sections do not need to contain only the changing action or scene description. For complex scenes, multiple characters, or longer chunks such as 10–15 seconds, it can be useful to repeat important details about the subjects, appearance, location, camera, and current scene state inside each section. More self-contained chunk prompts may sometimes give H3 stronger continuity and reduce cut-like transitions or drift.
Continuum carries visual/audio context from the previous chunk, but that does not guarantee that H3 will preserve every semantic detail automatically. Instead of relying only on phrases such as "continue from the previous chunk," describe the intended next action and important continuity details explicitly.
For example:
~~~text
Subject: A woman in a red jacket.
Style: cinematic nighttime photography.
[0-15s]
A woman in a red jacket stands at the nightclub bar, speaking with two people beside her. Medium shot, slow handheld camera movement, crowded dance floor visible behind them.
[15-30s]
The same woman in the same red jacket continues the conversation at the same bar with the same two people. The camera remains close and slowly moves around the group as one person replies.
[30-45s]
The woman turns toward the dance floor while remaining beside the bar. Keep the same clothing, nightclub environment, nearby characters, and continuous camera style.
~~~
The global preamble is optional. If fully self-contained prompts inside every Timeline section give better continuity for a particular model or scene, that is also a valid approach.
### Other prompt formats
**List** prompts use `---` separators.
**Fixed** reuses one prompt for every chunk.
**Auto** detects Timeline, List, or Fixed formatting from the supplied text.
If Timeline syntax cannot be parsed, V3.4 reports a warning and falls back to an applicable prompt mode rather than rejecting an otherwise runnable workflow.
## Run Storage
Run Storage preserves completed chunks and can reuse them after restarting ComfyUI. Regenerate From selects the first chunk to regenerate; earlier compatible chunks remain unchanged.
![Regenerate From](images/v34-regenerate-from.png)
### Keep the first part and regenerate the second part
For a normal T2VA or I2VA run configured as `2 chunks x 5 seconds`, Chunk 1 is the first five seconds and Chunk 2 is the second five seconds.
Before the first generation:
1. Set `Run Storage` to `Save + Auto Resume`.
2. Choose a fixed `Run Name` and keep the Base Seed fixed.
3. Generate the complete ten-second video once. Continuum stores the raw Video/Audio latent for each completed chunk.
To generate a different second part while keeping the first part:
1. Keep the same Run Name, Base Seed, model, LoRA, resolution, sampler, SIGMAS/steps, references, and Continuation settings.
2. Set `Regenerate From` to `Chunk 2`.
3. Leave `Variation Nonce = 0` for automatic new variations after completed regenerations, or choose a different positive value for an explicit variation. A compatible interrupted regeneration at `0` resumes its existing variation; reusing the same positive value does not request a different result.
4. Queue the workflow again.
Continuum loads the stored Chunk 1 instead of Sampling it again, uses its finalized end as the continuation context, and Samples a new Chunk 2. The report should show `1 reused, 1 generated`. The complete output is then decoded and assembled again; the saved first chunk is reused, while boundary Seam processing is evaluated during the new assembly.
If the two parts need different instructions, define them with List or Timeline Prompt syntax before the first generation. Changing the Prompt Plan or other generation-contract inputs after a stored run can create a new revision, in which case the previous Chunk 1 may not be reusable. A run originally generated with `Run Storage = Off` cannot be reused retroactively. `Regenerate From` works at chunk boundaries; it does not edit an arbitrary region inside one chunk.
FL2VA Long Terminal Merge is an important exception: its final two logical chunks are one atomic physical Sampling unit, so those two chunks are regenerated together.
- Use a fixed seed for reproducible resume.
- Changing persistent models, LoRAs, references, prompts, resolution, or sampling settings can create a new revision.
- Unknown upstream node classes no longer cause rejection by themselves; reuse still depends on compatible observable contracts and hashes.
## Spectrum interoperability
Spectrum is optional. Current Spectrum releases officially support H3 Continuum Interop API v1.
- Chunk 1 uses the normal initial path.
- Chunk 2 and later request Actual Prefix 2.
- Successful continuation logs accepted H3 Continuum API v1, actual prefix=2.
- Unsupported Spectrum versions fall back without a private patch.
See [Spectrum v0.2.15 H3 Continuum interoperability](https://github.com/xmarre/ComfyUI-Spectrum-MiniMax-H3#v0215-h3-continuum-interoperability).
Turbo LoRA and Spectrum are not mutually exclusive. Quality and speed remain workflow-dependent.
## Issue and pull-request response
### Issue #3
V3.4 removes blanket rejection of unknown upstream/custom class names. Run Storage evaluates the observable generation contract instead of treating an unfamiliar wrapper as automatically incompatible.
See [Issue #3](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/issues/3).
### Issue #4
V3.4 follows the current Core H3 layout contract and no longer requires the legacy frame_count parameter. The old strict-compatibility toggle is not part of the public interface.
See [Issue #4](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/issues/4).
### Issue #7
V3.4 allows First Frame, Last Frame, and Reference Images to be used together without requiring the separate BigStationW node or introducing a model allowlist. Hybrid presentation ordering and Run Storage identity are handled by Continuum while pure FL2VA and pure Ref2VA behavior remain unchanged.
The combined path was exercised locally with the B2049 hybrid variant. FL-only models were less consistent for this specific combination, so the README recommends a hybrid-capable model rather than rejecting other models in code.
See [Issue #7](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/issues/7).
### Pull request #1 and #2
These older partial pull requests were superseded by the consolidated [Pull request #5](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/pull/5). V3.4 selectively includes the compatibility direction needed for the current Core H3 contract, while the broader upstream synchronization and release-automation scope remains a separate upstream review.
The older pull requests are not represented as merged V3.4 changes.
### Pull request #5
PR #5 is the consolidated upstream-to-fork proposal covering current H3 compatibility, reference-input handling, assembly memory behavior, exact-duration handling, continuation-reference interoperability, and CI/release automation. V3.4 already contains the user-facing stability decisions described above; the PR remains an upstream review item and is not claimed as merged here.
## V3.4 input connection patterns
V3.4 separates the video-guide frame input from the driving-audio input. Choose the connection pattern that matches your source material.
### 1. Audio only
Connect `Load Audio` to `Driving Audio`. Use this when an existing song, dialogue track, or sound effect should remain the final audio. `Video Guide Frames` is not required.
![Driving Audio connection](images/v34-driving-audio-connection.png)
### 2. Video with its own audio
Connect `Load Video (Upload)` `IMAGE` to `Video Guide Frames`. If the uploaded video contains the audio you want to preserve, connect its `AUDIO` output to `Driving Audio` as well.
![Video Guide Frames and embedded Driving Audio connection](images/v34-video-reference-with-audio.png)
### 3. Video and audio from separate sources
Connect `Load Video (Upload)` `IMAGE` to `Video Guide Frames`, then connect a separate `Load Audio` node to `Driving Audio`. Use this when the video guide and the final audio source are different files.
![Separate Video Guide Frames and Driving Audio connection](images/v34-video-reference-separate-audio.png)
Both inputs are optional. Connect `Video Guide Frames` when video guidance is needed, and connect `Driving Audio` when the supplied audio should be preserved in the final output.
### Video Guide Frames frame rate
Use a 24 fps source for `Video Guide Frames`. `Load Video (Upload)` may accept files recorded at 25 fps or another frame rate, but acceptance alone does not guarantee correct temporal alignment with H3. For a non-24 fps source, set `force_rate` to `24` in `Load Video (Upload)`, or convert the file to 24 fps before loading it. If the source is already 24 fps, leave `force_rate` at its default and do not resample it.
## Current validation status
The public-surface suite checks ten V3.8X2 IDs and twelve IDs including V3.9, official workflows and matching ZIP payloads, declared external dependencies, Registry exclusions, preserved V3.8 widget/socket order, and presentation-only `Show Advanced Settings` / `Hide Advanced Settings` behavior. Registry payload hashes are recorded in `REGISTRY_MANIFEST.sha256`; source integrity is recorded in `MANIFEST.sha256`.
**Historical V3.8X2 preparation evidence:** the full CPU suite passed **1,367 passed / 1 skipped / 0 failed**. The prior dedicated GPU Functional Gate passed a `3 × 5 second` Review Each Chunk run: Q1–Q3 generated one physical group at a time, and Q4 reused all three groups with `3 reused / 0 generated`. Q3 and Q4 had identical decoded RGB and PCM SHA-256 values. The separate nine-reference Gate is recorded below. These checks confirm functional execution, prefix reuse, AV reconstruction, and the tested nine-reference path; they are not a full subjective image- or audio-quality rating.
### Nine-reference input gate (0.3 MP reference images)
On the tested NVIDIA GeForce RTX 5060 Ti 16 GB / 64 GB RAM system, a `3 × 5 second` Full Run completed with Reference Images 1–9, `512 × 608` output, Spectrum Off, LoRA Off, Sage Attention, Balanced 22-frame continuity, and test reference images sized to about `0.30 MP` each. Sampling used approximately `15.0–15.5 GiB` of the available `16,311 MiB` VRAM, so this is a high-resource configuration with little headroom on a 16 GB GPU.
Use the nine-reference bundle when the extra appearance or pose evidence is needed and the system has enough VRAM headroom. This single-seed gate confirms execution and basic visual continuity; it does not establish that nine references are universally higher quality than three, nor guarantee operation with higher-resolution references, larger output sizes, different models, or other concurrent GPU workloads.
For nine-reference use on a 16 GB GPU, pre-resize copies of the source images to around 0.30 MP before loading them. Internal conditioning resize remains active, but it should not be treated as a substitute for controlling the memory pressure of nine large source images.
Validation results apply to the tested local source and environment. They do not guarantee that an installed copy is current, that all models/wrappers fit every GPU, or that a different prompt will have the same visual quality.
## Limits
- V3.8 Second Pass / `refine_context` is not supported while a Review sequence is partial. V3.9 supports validated output-group conditioning for partial Review as described above.
- The 16GB GPU Gate passed with small VRAM headroom; memory use depends on the complete environment and workflow.
- Continuation does not guarantee frame-perfect identity or motion.
- Video Guide Frames guides H3; it does not reproduce every source frame.
- Driving Audio preserves selected audio, but visual lip synchronization remains model-dependent.
- FL2VA may settle on the supplied Last Frame before the requested duration ends; equivalent Core runs can show the same behavior.
- Long, high-resolution sequences may exceed system RAM during final decode and assembly even when chunk sampling succeeds.
- Disk-backed assembly lowers anonymous/private memory commitment for the final Continuum IMAGE, but Core Decode and downstream nodes may still allocate large RAM copies.
- Main Hi-Res Fix 3x5 2x is not accepted on the tested RTX 5060 Ti 16 GiB: the 37T group completed at 1152x1152, but the terminal 77T group failed at its first Second Pass inference with CUDA OOM. Reference/Hybrid-specific 1x5 acceptance passed; longer Hybrid/Reference cases remain unverified.
- Direct high-ratio interpolation with Latent Resize can create persistent H3 artifacts; use the Pixel/VAE Main path or an appropriate external H3 latent processor.
- Match Output can be substantially slower than 0.4 MP or 0.6 MP.
- Seam correction may keep the native boundary when a proposed correction is not safer.
- Optional template nodes must be installed, replaced, or bypassed by the user.
## License
See [LICENSE](../LICENSE).
+517
View File
@@ -0,0 +1,517 @@
# H3 Continuum — 日本語の更新履歴・実装資料
[現在の日本語README](../README_JA.md) · [日本語の詳細ガイド](USER_GUIDE_JA.md) · [CHANGELOG](../CHANGELOG.md)
正本`84899c1`(パッケージ3.9.8)の日本語READMEにあった更新説明・旧版画像・検証記録を保持しています。
「現行」「main上」などの表現、古い導入コマンド、版ごとの受入結果は当時の文脈です。
現在の導入案内や、未検証の条件を含む包括的な合格として読み替えないでください。
最新版の入手先と対応範囲は、日本語READMEを参照してください。
<!-- Preserved README sections begin. -->
**最新版パッケージは3.9.8です。** 固定版のGitHub配布は下表を参照してください。Registryの公開状態は別に確認します。
| バージョン | Release | カスタムノード本体ZIP | SHA-256チェックサム |
| --- | --- | --- | --- |
| 3.9.8 | [Release](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/tag/v3.9.8) | [ZIP](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/download/v3.9.8/ComfyUI-H3-Continuum-v3.9.8.zip) | [SHA256SUMS.txt](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/download/v3.9.8/SHA256SUMS.txt) |
| 3.9.7 | [Release](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/tag/v3.9.7) | [ZIP](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/download/v3.9.7/ComfyUI-H3-Continuum-v3.9.7.zip) | [SHA256SUMS.txt](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/download/v3.9.7/SHA256SUMS.txt) |
| 3.9.6 | [Release](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/tag/v3.9.6) | [ZIP](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/download/v3.9.6/ComfyUI-H3-Continuum-v3.9.6.zip) | [SHA256SUMS.txt](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/download/v3.9.6/SHA256SUMS.txt) |
| 3.9.5 | [Release](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/tag/v3.9.5) | [ZIP](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/download/v3.9.5/ComfyUI-H3-Continuum-v3.9.5.zip) | [SHA256SUMS.txt](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/download/v3.9.5/SHA256SUMS.txt) |
| 3.9.4 | [Release](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/tag/v3.9.4) | [ZIP](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/download/v3.9.4/ComfyUI-H3-Continuum-v3.9.4.zip) | [SHA256SUMS.txt](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/download/v3.9.4/SHA256SUMS.txt) |
| 3.9.3 | [Release](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/tag/v3.9.3) | [ZIP](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/download/v3.9.3/ComfyUI-H3-Continuum-v3.9.3.zip) | [SHA256SUMS.txt](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/download/v3.9.3/SHA256SUMS.txt) |
| 3.9.2 | [Release](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/tag/v3.9.2) | [ZIP](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/download/v3.9.2/ComfyUI-H3-Continuum-v3.9.2.zip) | [SHA256SUMS.txt](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases/download/v3.9.2/SHA256SUMS.txt) |
## 3.9.8 — KJ Model Preview Overrideの継続修正
監査済みKJ Model Preview Overrideでは、プレビュー画像を保持しても新Runの保存署名を変えず、別QueueでChunk 1を再利用してChunk 2だけ生成できます。対応は固定したKJソース・Core helperと、**VAE/Audio VAE未接続・Tiny VAE `none`**、Core NoPreviews/Auto/Latent2RGBに限定します。KJNodes 1.5.2全般への対応ではありません。更新後はComfyUIを再起動し、ブラウザーを再読込して新Runを作成してください。旧Take・Workflowの削除は不要です。[対応条件・監査ハッシュ・検証範囲](KJ_PREVIEW_OVERRIDE_V398.md)を参照してください。
### 公開版v3.9.8のGPU再検証 — 2026-10-16:PASS
通常ComfyUIに公開タグ **`v3.9.8`/`58fbd855`** を反映して再起動し、公開版464ファイルとの一致、実ログの3.9.8読込、新しいSampling v6契約内のKJアダプターと固定監査プロファイルを確認しました。新Runを2つ使用した計4 Queueの結果です。
| KJ Preview Override併用構成 | 別QueueのChunk 1→2継続 |
| --- | --- |
| Model Attention Backend:PyTorch/Spectrum OFF | **PASS — 1再利用/1生成** |
| Model Attention Backend:Comfy Kitchen INT8+監査済みローカルSpectrum | **PASS — 1再利用/1生成** |
両構成の2回目は画像を保持したKJ MODELをキャッシュ再利用(**8→16枚**)し、Samplerは実行。契約・revision・元Chunk 1 latentのSHA-256を維持し、`resume_safe=true`でした。映像・音声4本は全デコード成功(512×512/24fps、5秒120フレーム→10秒240フレーム)。保護対象170ファイル(保存Workflow 71件・既存Run/Take 46ファイルを含む)は全ハッシュ不変です。
**対応範囲:** 固定したKJ/Core監査プロファイルのみ。ローカルKJソースはupstreamと異なり、KJNodes 1.5.2全般の合格ではありません。Preview OverrideのVAE/Audio VAE未接続・Tiny VAE `none`・Core Auto、RTX 5060 Ti 16 GB、8-step Turbo/Euler/simple、2×5秒。Spectrumは監査済みローカル版のoffline replay、履歴・archiveはsystem RAM、model-aware OFFです。**未確認:** ブラウザーRunボタン/保存再読込、報告者の完全一致graph、別KJ/Spectrum版、SolAttn/preview VAE、カーネル単位の戻り値、主観画質・音質、速度・VRAM比較。保持PNGは32×32で、高解像度プレビュー品質の証明ではありません。既知Inspector問題は残り、旧unsafe/v5 Takeを自動修復・昇格しません。[Issue #31への結果返信](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/issues/31#issuecomment-6209571261)。
## 3.9.7 — 外部入力接続時の再生成キャッシュ修正
ChunksやVariation Nonceを外部ノードから接続した際に、明示したFull Run再生成がCoreのキャッシュで省略される問題を修正しました。前の採用済みChunk、既存Take、公開入力、公式Workflowは維持します。更新後はComfyUIを再起動し、ブラウザーを再読込してください。Workflowの変更や履歴の削除は不要です。 [Release notes](RELEASE_NOTES_V397.md).
**公開版3.9.7でのGPU再検証(2026-10-15):PASS。** 公開タグ(`18df80d`)で再起動後、外部Core INTからChunks=2・Nonce=0を入力し、完全に同一のプロンプトでChunk 2を2回再生成しました。自動Nonceは**1→2**へ進み、両回ともChunk 1を再利用してChunk 2だけを生成。以前のTake・raw latentも保持しました。基準生成を含む動画3本は240フレーム・映像/音声各10秒で全デコード成功。既存Workflow 60件・設定1件・Run/Take 30ファイル・公開版459ファイルのハッシュは不変です。条件はRTX 5060 Ti 16 GB、512×512、24fps、2×5秒、8-step Turbo、PyTorch attention/Spectrum OFF、Core API経由。CUDA/OOM/実行エラーは確認されず、ブラウザー・主観画質・他アクセラレーターの受入範囲は拡張していません。
元の[3.9.2 Source ZIP](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/archive/refs/tags/v3.9.2.zip)は従来のタグから取得できます。本体ZIPと公式Workflow ZIPは別の配布物です。
## 3.9.6 — 再生成指定・Nonce保持と自動Variationの再実行
V3.8/V3.9の**Regenerate From・Variation Nonceを保存・再読込後も保持**します。保存ONで**Full Run+明示Chunk+Nonce 0**を指定すると、QueueごとにRun Storageが処理し、完了後の再実行で新Variationを作成できます。前のChunkと既存Takeは保持します。更新後はComfyUIを再起動し、ブラウザーを再読込してください。旧markerなしChunk 1/Nonce 0は、Chunk 1を選び直して一度保存してください。[操作・互換性・検証範囲](REGENERATE_INTENT_V396.md)を参照してください。
## 3.9.5 — Issue #31:Attention/Spectrumの保存Take継続
CoreのModel Attention Backend(PyTorch/Comfy Kitchen INT8)と監査済みSpectrum Bindingを正確に識別し、新しいRunのReview Each Chunk継続を可能にします。公式Spectrum v0.2.29も専用プロファイルで対応します。以前のunsafe履歴は保持し、再利用可能へ自動変換しません。更新後は新しいRunで生成してください。公開ノード・Workflow・Sampling v6は維持します。[対応実装・更新手順・検証範囲](ISSUE31_SAMPLING_SIGNATURES.md)を参照してください。
## 3.9.4 — Registry配布の梱包修正(2026-10-14)
Manifestに記載したReview復元ガイドを収録し、配布対象外のプロンプト作成資料を除外します。公開前に実際のRegistry ZIPを作り、全収録ファイル・CRC・Manifestの全ハッシュを照合します。生成動作・12公開ノード・公式Workflow・保存済みRun/Takeは変更しません。3.9.3と3.9.2のGitHub Releaseは保持します。
## 3.9.3 — 一時ストレージの保存先設定(2026-10-13)
未圧縮フレームなどの一時データを、指定したストレージへ保存できるようになりました。**設定 → H3Continuum → Storage → 一時ストレージを管理**を開き、**ComfyUIサーバー側**の絶対パスを入力するかストレージ候補を選択して、検証・保存してください。空欄なら従来の保存先を使用します。ワークフローの変更は不要で、V3.9ノードと3.9.1/3.9.2のワークフローを引き続き使用できます。
- **FinalizeとDecode Cache Helperの両方**が対象です。指定先の`h3-continuum-scratch`専用領域へ保存します。Run/Takeや完成動画の保存先は別です。
- 現在・以前の保存先、一時ファイル量、空き容量、使用中/削除可能な容量を表示します。**未使用の一時ファイルを削除**は、参照中の出力・Tensorのview・生存プロセス・所有者不明のファイルを保護します。キャッシュ削除後は再デコード・再書き込みが必要になる場合があります。
- 保存・削除は実行中と待機中のQueueが空のときに行います。新しいファイルから保存先を切り替え、既存出力は元の場所で保持します。指定先が使えない場合、元のSSDへ黙って戻しません。マルチユーザー運用では共有保存先のUIは閲覧のみです。
- 使用方式は既存ノードで選びます。**Finalize:Auto/RAM/Disk-backed、Decode Cache Helper:Auto/RAM/Off**。Autoの4 GiB境界とRAM安全余裕は維持します。手動RAMは確保前に空きメモリを、手動Disk-backedは空き容量を確認します。ただし、実行中のメモリ・容量変動やドライブ切断まで防ぐ保証はありません。VRAMを保存領域にする機能は追加していません。
**保存先変更は書き込み先を変える機能で、総書き込み量を減らすものではありません。** 物理ストレージが1台なら、同じ機器の別フォルダー・別パーティションへ移しても書き込みを逃がせません。別のSSDは速度を保ちやすく、HDDも使用できますがDecode・結合・保存が遅くなる場合があります。十分なCPU RAMがあればRAM方式で一時書き込みを避けられますが、余裕がなければページファイルへの書き込みやOOMが起こり得ます。CacheをOffにするだけではFinalizeのDisk-backedを無効にしません。容量表示は論理ファイルサイズであり、物理占有量やSSDの累計書き込み量ではありません。対応先はローカルファイルシステムで、ネットワーク共有は正式対応外です。
更新時はカスタムノードを更新し、ComfyUIを再起動してブラウザーを強制再読込してください。その後の保存先設定の変更には再起動は不要です。空欄時はFinalizeがComfyUIのtemp、Decode Cacheが既存の`H3_DECODE_CACHE_TEMP_DIR`指定またはOSのtempを使います。フォルダーを指定した場合、その指定がキャッシュ用環境変数より優先します。
映像のdtype、動画・音声の結合計算、生成条件の識別は維持します。[実装仕様](TEMPORARY_STORAGE_JA.md)と[検証記録](../PACKAGE_VALIDATION.txt)も参照してください。設定画面のスクリーンショットは[英語README](../README.md)の冒頭にあります。Releaseの公開状態は上表を参照してください。Registry公開は別に確認します。
画面にHTTP 404が出る場合は、ComfyUI起動ログを確認してください。旧プロセスが同じアドレス・ポートを占有していると、更新後の起動が失敗し、新しい画面だけが表示される場合があります。残っている旧環境を終了するか別のポートを選んで、新しいComfyUIを起動してください。**Refresh(更新)**は保存先の状態を再取得するボタンで、ComfyUIの再起動ではありません。取得失敗時はHTTP結果と対応方法を表示し、状態を取得できるまで保存・削除を無効にします。
## 3.9.2 — Issue #30 Frontend hotfix(2026-10-13)
ReviewとReference Imagesの不要なUI処理を削減します。
- Review比較基準がない場合、上流設定snapshotの作成を省略。
- Reference選択時のSampler通知を1回へ整理。native値の即時更新と元widget callbackを維持。
- Referenceノードの目標寸法と実寸が同じ場合、不要なサイズ変更を省略。
比較基準がある場合の監視と、Take/Queue時の最新設定確認を維持します。Sampling、保存済みRun/Take、公開Node ID、公式Workflowのbytesは変更しません。引き続きV3.9ノードを使用してください。
Frontend修正で**CPU/JavaScript回帰1703件成功・2件スキップ**、独立したInspector再configure問題の**2条件を期待失敗として記録**しました。Coreのオフライン登録・計画確認と隔離Chromium検証も通過しています。このパッケージの実ComfyUIブラウザー確認は[PACKAGE_VALIDATION.txt](../PACKAGE_VALIDATION.txt)に記録します。モバイルでのフリーズ解消は未確認です。Reference表の再構築と、比較基準がある場合のReview走査は残ります。今回のhotfixで新たなGPU生成は行っていません。
更新後は**ブラウザーを強制再読込(Ctrl+F5)**してください。表示するパッケージ番号も更新する場合は、通常の手順でComfyUIを再起動してください。3.9.1のRun/Takeは、このUI修正のために移行する必要はありません。固定版は[GitHub Releases](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases)を参照してください。Registry公開は別作業です。
## 3.9.1で変わった点(2026-10-04)
Second Pass、Takeの再利用、Review音声に関するバグを修正しました。
- **Second Pass**:音声のみ/複合キーフレームを保持し、途中Reviewや中間Chunk単独でも、その出力区間のconditioningとReference割当を照合して引き継ぎます。
- **Takeの再利用**:Samplerのclosure、MODELのCFG/wrapperも互換性判定に含め、異なる生成設定のTakeを誤って再利用しないようにしました。
- **ReviewのDriving Audio**:物理groupの自然時間に合わせて元PCMを1回だけ切り出します。Exact ON/OFFで開始位置を揃えます。
- **音声resample**:ComfyUI Core標準APIを優先し、古いCoreでは従来方式へ戻します。同じVAE内部の重みを直接変更する場合のReference Encode Cacheの制約も下記に明記しました。
**3.9.0から更新する場合**:更新後にComfyUIを再起動し、ブラウザーも更新してください。旧v5 Takeと履歴は読めて、削除されませんが、新しいv6生成の前半として再利用できません。新規Full Video、または **Start again from Chunk 1** で開始してください。公開Node ID・widget順・公式Workflowは維持し、引き続き **V3.9ノード** を使います。
修正済みコードは **CPU 1,696成功・1スキップ**、対象を絞ったReview Second PassのGPU検証、生成/Take再利用の統合GPU検証に合格しています。ブラウザー受入は未完了です。任意のLATENT拡大で出る二重輪郭は別課題の [Issue #27](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/issues/27) として管理し、今回の修正の必須Gateには含めません。今回もmainソースの更新で、GitHub Release/Registry公開ではありません。
![H3 Continuum V3.9:固定の参照画像番号、チャンク別割り当て、Samplerへの1本の接続](images/v39/v39-feature-summary.png)
*現在の`main`はV3.9です。Image 1~9の固定番号、チャンク別のReference選択、V3.9 Samplerへの1本の接続が新しい操作です。*
## 3.9.1の修正詳細
Second Passで音声のみのキーフレームを保持し、複合キーフレームでは映像だけを空間適応します。ReviewのDriving Audioは、区間選択前の物理decode groupの自然時間から元PCMを切り出します。Exact ON/OFFの開始位置は共通です。完全に空のReview音声だけは選択映像の尺に合う無音を補い、部分的に短い音声はExact OFFで引き延ばしません。
新しい生成はSampling Contract v6/Graph Contract v4を使用します。Samplerのコード・closureとMODELのCFG/wrapper等の設定も再利用判定へ含めます。旧v5のTake・raw・生成履歴は読めますが、v6生成へ継ぎ足せません。新規Full Video、またはReviewの「Start again from Chunk 1」で新しいrevisionを作成してください。旧Takeは残ります。署名を観測できない外部設定でも新規生成は可能ですが、自動再利用は無効になります。明示した不適合TakeやRetryを別Takeへ置き換えません。
Reference/DrivingのresampleはCore標準APIを優先し、そのAPIがない旧Coreだけ従来のTorchAudioを使います。最終出力へ渡す元PCMの波形とsample rateは保持します。
**Reference Encode Cacheの制約**:同じVAEオブジェクトの内部重みを直接書き換えた場合、その変更を自動検知しません。変更後は`v3.ref_encode_cache.clear_ref_encode_cache()`を呼ぶか、新しいVAEオブジェクトへ再ロードしてください。通常の同一VAE・同一入力ではcache HITを維持します。Decode Cache Helperのresetは別のcacheを対象とし、Reference Encode Cacheを消しません。
この修正はV3.9内の変更です。公開Node ID・socket・widget順、公式Workflow、Run Storage v3、State/Session形式を維持し、LoRA Planは追加しません。CPU検証と実ブラウザー・GPU受入は別です。
V3.9では、Second Passへ渡す条件を実際のReview出力区間へ対応させます。途中までの出力や中間Chunk単独でも、条件の照合に成功すれば、その区間自身のconditioningとReference割当を継承します。全体の完了状態や保存済みTakeは変更しません。条件が不足・不一致の場合は、従来の診断付きprompt-only fallbackを維持します。
修正済みランタイムは**CPU 1696成功・1スキップ**、**512×512・24fpsのGPU 3ケース**(途中2区間、Chunk 2単独、完了済み3区間の再利用)で合格しました。全区間の条件継承を確認し、保存FLACと元PCMの対応区間は完全一致、既存Takeとrawファイルも保持しています。合格対象は修正対象の動作です。ブラウザー受入は未完了で、任意のLATENT拡大時の二重輪郭はIssue #27で別途調査します。過去のV3.8 Release/tagはそのまま利用できます。
## V3.9の基本操作を動画で見る
[H3 Continuum V3.9 — Getting Started: Chunks, Prompts & Reference Images](https://youtu.be/AKJxBXaiG6Q)(12分26秒)
First Imageの基本設定、チャンクの設定、Review/Takeでの再試行、Timeline形式のプロンプト、Reference Imagesの使い方を紹介します。英語のナレーションと字幕、7つのチャプターがあり、バージョンを明記したUIの静止画と実際の生成例で説明します。
![公開当時のH3 Continuum V3.8X2機能紹介画像](images/v39/v38x2-feature-summary-historical.png)
*V3.8X2の画像も比較用に残しています。画像内の「CURRENT MAIN」は作成当時の表記で、現在の`main`はV3.9です。V3.8 SamplerとV3.8X2 Workflowは利用できますが、Reference配線はV3.9へ自動変換されません。*
## main更新:外部INTによるWidth/Height入力の修正(2026-10-01)
V3.8/V3.9のWidth/Height入力と表示コントロールの対応を修正しました。外部INTの接続位置をWorkflow保存・再読込やタブ切替後も維持します。接続中の欄は編集不可になり、表示する数値は保存済みのManual予備値です。上流ノードが実行時に出す値を推測して表示するものではありません。接続を外すと手動編集に戻ります。
Coreの **Int** ノードから正確な幅・高さを指定する場合は、**Size Source = Manual**にしてWidth/HeightのINT入力へ接続してください。**First Image**は引き続き画像とResolution設定からサイズを決めます。INTを接続しても優先順位やSize Sourceは自動で変わりません。有効なFirst Imageがない場合は従来のManualフォールバックを使います。寸法検証、Backend入力順序・初期値は変更していません。
確認済み:**全CPU試験1569 PASS/1 skipped**(新Frontend試験22/22・既存Review試験60/60を含む)、両バージョンのChrome保存・再読込とタブ5往復。外部INTを使ったV3.9のGPUテスト1回では、**480×640・24fps・96フレーム・映像/音声とも4秒**を正確に出力しました。20ステップ、全体**107.3秒**で、保存動画の全フレームを正常にデコードできました。短い単一条件の確認であり、長時間継続や全モデルでの受入確認ではありません。Sampling、Run Storage、先行するReview/キャッシュ修正、公式Workflow JSON/ZIP、過去Releaseは変更していません。
Continuumを`main`から更新し、**ブラウザーをCtrl+F5で強制更新**してください。このJavaScript修正だけならComfyUI本体の再起動は不要です。同時にPython修正や新ノード定義も更新された場合は通常どおり再起動してください。今回もmainのソース修正で、新Release/Registry公開ではありません。
## main更新:V3.9の条件付け準備の速度修正(2026-10-01)
**Prompt Format = Fixedかつ有効なReference Imagesがない場合**に、V3.9で余分なプロンプト条件付けの準備が繰り返される経路を修正しました。入力が同じ場合に既存の上限付きconditioningキャッシュを再利用します。First Imageは利用できます。List/Timeline、有効なReference Images、Reference Audio、Timeline Video、終端ペアのgroupは従来の経路を維持します。Decode Cache Helperとは別の修正で、Samplingは省略しません。
検証した`2 × 5秒`・480×640・First Imageの条件では、group準備が約**10.76秒から0.013秒**へ短縮しました。修正後の同セッション比較(warm-up後、各版3回)では、全体時間の中央値は**V3.8が273.58秒、V3.9が273.78秒**で測定範囲も重なり、以前の約6%差は見られなくなりました。全Workflow/モデル/GPUでの解決や、GPU演算そのものの高速化を確認したという意味ではありません。
Continuumを`main`から更新し、Python修正を読み込むため**ComfyUIを完全に再起動**してください。確認済み:**CPU 1558 PASS/1 skipped、GPU生成8回すべて成功**(ComfyUI 0.38.0/RTX 5060 Ti 16 GB)。Seed、チャンク/フレーム契約、Run Storage、V3.8X2/V3.9公式Workflow・ZIP、過去Releaseは変更していません。今回もmainのソース修正で、新Release/Registry公開ではありません。
## main更新:Review表示の復元修正(2026-10-01)
別のWorkflowタブへ切り替えて戻ると、保存済みTakeが残っていてもReview操作が消える問題を修正しました。Workflowの復元完了後に履歴を取得し、古い画面の非同期結果は反映しません。実際に生成設定を変えた場合のReview/Take適用ガードは維持し、元の設定に戻すと操作が復帰します。
カスタムノードを`main`から更新後、ブラウザーを更新して新しいJavaScriptを読み込んでください。確認済み:Frontend回帰60/60、関連CPU試験108件、Chromeのタブ5往復とBase Seed変更/復元。Take管理データは不変です。Sampling、Run Storage契約、V3.8X2/V3.9のWorkflowファイルと過去Releaseは変更していません。この修正のGPU継続生成は未検証です。[詳細と検証範囲](REVIEW_RESTORATION_REPAIR.md)。
## アーカイブされた実装・受入記録(V3.8現行ガイドではありません)
以下は一時的にsource historyとして残る旧リリース記録です。V3.8の標準経路・公開範囲・推奨配線を示すものではありません。旧WorkflowにはMigration Policyと対応するhistorical Release/tagを使用してください。
### V3.7 高解像度Refinement基盤
V3.7では`H3 Continuum Sampler V3.7`を追加し、解像度を変更するSecond Passに必要な2つの基盤を完成させました。V3.6のProduction初期値と既存SIGMAS socketは変更していません。
### 解像度変更後もConditioningを再構築
Conditioning Adapterは、First/Last Image、Continuation context、Still Image Guideの元画像をtarget latent geometryに合わせて再構築します。Physical group ownership、Long Terminal Merge、Core側のPackedLayout/RoPE再構築、First Pass Audioのpassthroughは維持します。704×704と1152×1152のCorrectness GateはPASSしていますが、大きなcanvasほどmemory使用量と処理時間は増えます。
### RefineScheduleでRefinement範囲を明示
内部のversioned contractである`RefineSchedule`は、`External`、`Full`、`Tail`、`Partial`に対応します。`External`は入力SIGMAS tensorをそのまま維持し、`Tail`と`Partial`は同じsource scheduleから範囲を正確に切り出します。内部Tail 6 Gateでは、6 evaluationのsuffix再現とdecoded Audio PCMのbit-exact passthroughを確認しました。現時点では実装基盤として内部で扱い、公開widgetは追加していません。Production初期値も変更していません。
### Still Image GuideはExperimental
V3.7では、Still Image Guide 1枚をabsolute frameからowner physical groupへ割り当て、高解像度Second Passでも元画像から再encodeできます。位置変換、Terminal Merge ownership、Run Storage identity、non-owner groupの非干渉はCorrectness PASSです。一方、Core Add Guideはhard anchorとして働くため、anchor位置で急なtrajectory変化が起き、その後のmotionも別の軌道へ分岐する場合があります。**Still Image GuideはExperimentalで、Production昇格はHOLDです。** 滑らかなtransition制御としては扱わないでください。
### V3.6.1 Maintenance Hotfix
V3.6.1は、受入済みBalanced 22 Masked AV経路を変更せず、限定的なfail-open/互換性問題を修正します。Timeline/List記法が混在したPromptは`H3C-P105`を表示し、Timeline解析ではstandalone `---`行だけを除去します。未指定Chunkの既存fallbackとFixed Promptのfail-open動作は維持し、Prompt構文だけを理由に生成を停止しません。
`Continuation Backend = Standard`かつAudio Continuity ONでは、`Balanced — 22 frames`が引き続きMasked AVを使用します。`Fast — 5 frames`、`Strong — 39 frames`、`Auto — conservative`はRun Storage identity作成前に、受入済みReference Contextへ解決します。UIの選択値は書き換えず、status reportへ解決結果と理由を表示します。Audio Continuity OFFのStandardはMasked Video、Compatibilityは常にReference Contextのままです。
Run Storageは、Last Frame未接続を表す値をすべて同じ空identityとして扱います。通常のLast Frameなし2チャンク生成を3チャンクへ延長すると、旧最終Chunkが`none`だったV3.6.1 cacheを含めて`2 reused, 1 generated`になります。実Last Frame接続時とLong Terminal Mergeのatomic pairは従来どおり厳密です。Run StorageをOffへ変更した場合は、Queue前にhiddenのRegenerate FromとVariation Nonceも`Auto`と`0`へ戻します。
### V3.6 Masked AV Continuation
V3.6では`H3 Continuum Sampler V3.6`を追加しました。標準backendは、前Chunkで確定したVideo/Audio latent prefixを次のH3 Target内へ直接配置し、CoreのNoise Maskで保持します。V3.5のReference Context方式のように、同じ過去フレームを別Reference blockとして追加しないため、H3が処理するpacked sequenceを短縮できます。
| Continuation Backend | 用途 |
|---|---|
| `Standard` | 推奨V3.6経路。Balanced 22+Audio ContinuityはMasked AV、Fast 5/Strong 39/Autoは安全にReference Contextへfallbackします。Audio Continuity OFF時はMasked Videoです。 |
| `Compatibility` | 受入済みV3.5 Reference Context動作へ戻すAdvanced fallbackです。 |
内部transport識別子はUIへ表示しません。Run Storage identityは実行前に解決したtransportへ従います。Balanced 22 Masked AVはReference Contextと分離し、非Balanced Standard fallbackは同一動作のReference contractを安全に共有します。`Regenerate From`でもLong Terminal Mergeのlogical `[2,3]`を1つのphysical sampleとして再利用・再生成します。
### 受入結果
- T2VA、I2VA、Reference Image+Reference Audio、Balanced 22-frame Joint AV、3×5秒FL2VA Long Terminal MergeのGPU GateをPASS。
- 保護Video/Audio prefixは最終的にbit-exact。生成領域はSamplerの出力を維持します。
- Reference Image+Reference Audio代表出力は実聴PASSで、click、dropout、定位異常の報告なし。
- 実測したFL2VA Terminal Group 2では、StandardがCompatibilityよりpacked rowsを2,874行(`8.01%`)削減し、Sampling中央値を`4.06%`短縮。速度はmodel、hardware、workflowで変わります。
- `chunk_seconds`は4.0~30.0秒、初期値5.0秒、step 0.1です。5~15秒を推奨・検証済み範囲として維持し、長時間・高解像度ChunkはVRAMと処理時間が大きく増える場合があります。
V3.5.3、V3.5、V3.4のNode IDと保存済みWorkflowは維持します。V3.5.3は従来Reference Contextのままで、V3.6標準backendへ暗黙に切り替えません。
## V3.5.3 Maintenance Hotfix
V3.5.3は配布整合性だけを修正するメンテナンス版です。公開Conditioning Bridge WorkflowのWidth/Height接続、V3.4/V3.5テンプレート内の孤立link ID、Hybridの`<Picture N>`警告番号、厳密な画像decoderで読めなかった旧PNG 1件を修正しました。
生成動作、Node、socket、Sampling、Conditioning payload、Terminal Merge、Assembly、Seam、Run Storage、Prompt/CLIP cache、Video Guide最適化、V3.4互換性は変更していません。既存V3.5.x Workflowはそのまま読み込めます。
## V3.5.2 Stabilization & Optimization Update
V3.5.2は新しい生成モードを追加する版ではありません。V3.5.xを安定化し、安全な条件で繰り返しPrompt/CLIP処理を省略し、長いVideo Guide入力の一時メモリを削減します。既存Workflowと生成契約は維持します。
![H3 Continuum V3.5.2の安定化・最適化結果](images/v352-stabilization-optimization.png)
画像は受入済み最適化の実測値です。最終パッケージゲートでは、さらにComfyUI 0.33.3と、この画像を含む172項目のManifestでPASSを確認しています。
### 主な変更
1. **繰り返し実行Prompt/CLIP cache**: 条件が変わらないT2VA、I2VA、FL2VAではCPU上のPrompt/CLIP結果を再利用できます。Reference、schedule、tokenizer option、hook変更CLIPは安全側でbypassします。Sampling自体は通常どおり再実行します。
2. **Video Guideメモリ最適化**: source全体をfinite検査とSHA-256 identityへ含めたまま、VAE/conditioningに必要なprefixだけを保持します。
3. **安定化監査**: Sampling、Decode、Assembly、保存、任意機能の負荷、保持memoryを実測しました。根拠のないSampling、Driving Audio、Audio hash、Assembly、Seam、Session、V3.4最適化は採用していません。
### 実測結果
| 項目 | 変更前 | V3.5.2 | 結果 |
|---|---:|---:|---:|
| T2VA繰り返しPrompt/CLIP | 5.898398秒 | 0.000028秒 | Cache HIT、`encode_calls=0` |
| FL2VA繰り返しPrompt/CLIP | 21.642687秒 | 0.007128秒 | Cache HIT、`encode_calls=0` |
| Video Guide Peak追加RSS | 396.8 MiB | 114.7 MiB | 約71%削減 |
| Video Guide保持Storage | 225 MiB | 93 MiB | 約59%削減 |
Prompt/CLIPの数値はconditioning区間だけで、総生成時間ではありません。Video Guideメモリ比較は300×256×256 RGB入力/必要prefix 124 framesの固定CPU条件です。別のGPU A/Bでは実際の15秒/360-frame Video Guideを使い、decoded videoとaudio PCMの完全一致を確認しています。
RTX 5060 Ti 16 GB/RAM 64 GBの検証環境で測定したSage-only Production baselineは、576×576 T2VA 1×5秒が168.069秒、640×640 FL2VA Long Terminal Merge 3×5秒が379.765秒です。環境・設定固有の測定値であり、すべての環境に対する速度保証ではありません。Samplingが最大コストで、Continuum Assemble + Seamは1%未満でした。
**V3.8.0はhistorical release baselineです。V3.8X2はpackage 3.8.3の別Workflowとして維持し、V3.9はmainの現行ソースです。** V3.8X2が内部利用する旧module/classはsourceへ維持します。現在の登録面はV3.8X2の10 IDとV3.9追加の2 IDです。一部は旧IDを維持しますが、それ以外のIDを必要とする旧保存Workflowは対応するhistorical Release/tagを使用してください。Still Image Guideは引き続きExperimentalです。
## V3.5.1 Reference Audio/互換性更新
V3.5.1ではV3.4のSampling、Conditioning、Terminal Merge、Assembly、Seam、Run Storage契約を変更せず、次の2点を追加しました。
1. **Reference Audio(任意)**: 生成音声を置換しない、H3 nativeの音声conditioningです。Workflowの保存・再読込互換を守るため、2つのsocketは常時表示します。
2. **Conditioning Bridge V3.5**: Continuumのphysical groupごとに、完全なCore互換`MODEL`と`CONDITIONING`を外部Samplerへ公開します。
Reference Audio socketはPython `INPUT_TYPES`で常設し、保存・再読込時の値ずれを防ぐため動的socket変更とUI-onlyの`Hidden`/`Show` controlを削除しました。V3.4のNode IDとbackend socket keyは保存済みWorkflow互換のため維持しており、既存V3.4/V3.5 Workflowの接続は変わりません。
### Reference系入力の違い
表示名は役割の違いを示しています。名前が似ていても同じ組み合わせではありません。
| 表示入力 | 接続元 | 用途 | 最終音声 |
|---|---|---|---|
| `reference_image_1`~`reference_image_3` | 静止画`IMAGE` | 人物、被写体、外観などの継続参照 | 音声なし |
| `Video Guide Frames` | Video loaderの`IMAGE`フレームbatch | 全chunkでmotion、framing、timing、appearanceをガイド | 元動画の音声は含まない |
| `Driving Audio` + `Driving Audio VAE` | Audio loader、またはVideo loaderの`AUDIO`と対応Audio VAE | 音声で生成をガイドし、入力音声を最終出力へ維持 | 生成音声を選択した元音声へ置換 |
| `Reference Audio (Optional)` + `Reference Audio VAE (Optional)` | 単独Audioと対応Audio VAE | H3 native conditioning専用 | 生成音声をcopy・置換しない |
音声付き動画を使う通常の接続は、loaderの`IMAGE`を`Video Guide Frames`へ、`AUDIO`を`Driving Audio`へ接続します。conditioning専用動作が目的でない限り、動画の音声を`Reference Audio (Optional)`へ接続しないでください。
![Sampler V3.5の常設Reference Audio、Video Guide Frames、Driving Audio、Reference Image入力](images/v351-video-guide-frames.png)
`Reference Audio (Optional)`と`Reference Audio VAE (Optional)`は、Pythonのinput schemaどおり常時表示します。V3.5.1ではこの2 socketをFrontendで動的に追加・削除せず、UI-onlyの`Hidden`/`Show` widgetも使用しません。これによりWorkflowを保存・再読込した際のwidget値の位置ずれを防ぎます。backend input keyと既存接続は変更しません。
![常設Reference Audio socketと正常なwidget配置](images/v351-reference-audio-permanent.png)
### Conditioning Bridge V3.5
`H3 Continuum Conditioning Bridge V3.5`は外部sampling用のAdvanced接続点です。`model`、`clip`、外部処理後の`video_latents`、`assembly_plan`、`refine_context`を接続し、選択したconditioning経路で必要な場合だけ`video_vae`を接続します。出力`group_models`と`conditioning`はphysical group数と同数で、各要素は完全なComfyUI objectのままです。CONDITIONING内部entryをphysical-group listへflattenしません。
```text
H3 Continuum Sampler V3.5.video_latents
→ LBH等の外部H3 latent処理
→ H3 Continuum Conditioning Bridge V3.5
→ Core BasicGuider + 外部Sampler
→ Core Video / Audio VAE Decode
→ H3 Continuum Assemble + Seam V3.5
```
AV LATENTの組、Noise、SIGMAS、Audio Lock、sampling、audio passthroughは外部workflow側の責任です。Bridgeは準備済み`MODEL`と`CONDITIONING`、更新済みAssembly Planだけを公開します。
![V3.5.1 LBH+Conditioning Bridge外部sampling接続図](images/v351-lbh-conditioning-bridge-flow.svg)
退役したV3.5.1 LBH+Conditioning Bridgeの接続例は、対応するhistorical Release/tagでのみ参照できます。これはV3.8 Workflowではありません。
接続例ではCore標準名の`BasicGuider`、`BasicScheduler`、`SamplerCustomAdvanced`を変更していません。公開サンプルではCoreノードを独自タイトルへ変更せず、Continuumノードや外部カスタムノードと直感的に区別できる状態を維持します。LBH latent upscaler、AV LATENTの結合/分離、load/save、任意accelerationは外部ノードなので、各ComfyUI環境に合わせて導入または置換してください。
LBHはlatent geometryだけを変更し、Continuumのdenoise強度は持ちません。拡大後latentをどれだけ再生成するかは外部`BasicScheduler`のSIGMASで調整します。1.5x等へ上げても高速な場合がありますが、target canvasに応じてmemory、decode、sampling負荷は増加します。
## V3.5の主な追加機能
V3.5の大きな追加は2点です。
1. **Continuum対応Second Pass / Hi-Res Fix**: physical group構造を維持した再Samplingと、標準Pixel/VAE 2x経路。
2. **低メモリAssemble + Seam**: Exact Duration、Seam、Terminal Merge、Audioを維持したRAM / Disk-backed / Auto出力。
### 1. Second Pass / Hi-Res Fix
`H3 Continuum Hi-Res Fix V3.5`はVideo VAE Decode → pixel resize → Encode → low-denoise Second Passを1ノード化したMain経路です。長尺2xではVRAM上限に達する場合があるためExperimentalです。
![H3 Continuum Hi-Res Fix V3.5 node](images/v35-hires-fix-node.png)
通常のHi-Res Fix接続:
```text
H3 Continuum Sampler V3.5
→ H3 Continuum Hi-Res Fix V3.5
→ Core Video / Audio VAE Decode
→ H3 Continuum Assemble + Seam V3.5
```
`H3 Continuum Second Pass V3.5`は外部加工したH3 video latentをphysical group構造のまま再Samplingする安定版Advanced入口です。Issue #8への本体回答です。
![H3 Continuum Second Pass V3.5 node](images/v35-second-pass-node.png)
外部latent upscaler接続:
```text
H3 Continuum Sampler V3.5.video_latents
→ LBH等の外部H3 latent処理
→ H3 Continuum Second Pass V3.5
→ Core Video / Audio VAE Decode
→ H3 Continuum Assemble + Seam V3.5
```
画像では`BasicScheduler`を省略していますが、Second Pass用`SIGMAS`の接続は必要です。GPU受入済み基準は`res_multistep`、`simple / 10 steps / denoise 0.35`、Lanczos pixel resize、FL2VA 1×5秒・576→1152です。Hybrid FL2VA + Referenceも1×5秒で、統合2x Hi-Res FixとAdvanced Second Pass直結の両方を受入しました。
Hi-Res Fixを接続しなければV3.4 Sampling経路は変わりません。強制CPU offloadや低VRAM samplerは追加していません。`enabled=false`ではResize、VAE、Second Passを評価せず、元のVideo/Audio LATENTとAssembly Planをそのまま返します。
### 2. 低メモリAssemble + Seam
`H3 Continuum Assemble + Seam V3.5`はRAM / Disk-backed / Autoを選択できます。Disk-backedでは完成Video IMAGEをWindows対応mapped fileへ直接writeし、AudioはRAMに維持します。
1536×1536・360 frames・完成IMAGE 9.49 GiBの実測では、RAM版とDisk-backed版の映像/音声hashが完全一致しました。
| Windows process指標 | RAM | Disk-backed | 削減 |
|---|---:|---:|---:|
| Private memory | 12.41 GiB | 2.91 GiB | **9.50 GiB** |
| USS | 10.29 GiB | 0.80 GiB | **9.49 GiB** |
これはContinuum Assembly時の**システムRAM/private commitment削減**であり、SamplerのGPU VRAM削減値ではありません。Disk I/Oにより最終Assemblyが遅くなる可能性はありますが、前段のモデルSampling速度は変えません。
### GPU受入範囲
| 経路 | 条件 | 結果 |
|---|---|---|
| Main Hi-Res Fix | FL2VA 1×5秒、576→1152 | PASS |
| Main Hi-Res Fix | Hybrid FL2VA + Reference 1×5秒、576→1152 | PASS |
| Advanced Second Pass | Hybrid/Reference 1×5秒、576×576 | PASS |
| Main Hi-Res Fix | FL2VA 3×5秒、576→1152、RTX 5060 Ti 16 GiB | **Terminal 77TでCUDA OOM** |
3×5秒2xではFirst Passと37T groupのSecond Passは完了し、最後のlogical `[2,3]`を維持した77T physical groupが1152×1152の最初の推論でVRAM上限に達しました。group構造の破損ではなく、16 GiB実機でのリソース上限です。
詳しいsocket接続・OFF時のpassthrough・既知の制限は[英語版V3.5接続ガイド](V35_HIRES_FIX.md)を参照してください。
## 公開テンプレートワークフロー
- [V3.9 Workflow JSON](../examples/workflows/MiniMax_H3_Continuum_V39.json)/[単独ZIP](../examples/workflows/MiniMax_H3_Continuum_V39.zip) — 現行V3.9 graph。ZIPにはこのJSONだけを収録
- [V3.8X2 Workflow JSON](../examples/workflows/MiniMax_H3_Continuum_V38X2.json) — 併存する旧系公式graph。Spectrumは初期状態でOFF
- [V3.8X2 Workflow ZIP](../examples/workflows/MiniMax_H3_Continuum_V38X2.zip) — V3.8X2とV3.9の両JSONを格納。カスタムノードのインストーラーではありません
- [V3.8X2 Helper明示名JSON](../examples/workflows/MiniMax_H3_Continuum_V38X2+Decode_Cache_Helper.json)/[ZIP](../examples/workflows/MiniMax_H3_Continuum_V38X2+Decode_Cache_Helper.zip) — 同じgraphを説明的な配布名で収録
Registry配布予定の対象はV3.9とV3.8X2のJSON/ZIPです。依存なしのテンプレートではなく、Spectrum・rgthree・KJNodesが必要です。ComfyUI-Easy-Useは不要です。Turboへ切り替えてもgraph内の外部ノードは残ります。SpectrumをOFFにし、用途に合うLightX2V Turbo LoRAを1本だけONにしてsampler/Stepsを合わせます。保存状態はSpectrum OFF、Turbo LoRA OFF、res_multistep/simple/20 Stepsです。
配布元のprompt、画像・音声名、ノード表示名、設定は変更していません。手元にあるファイルを選び、不要な入力をOFFにし、promptを入力してください。モデルやメディアは同梱しません。`Save 3x5s Video`などの保存済み表示名は生成時間を決めません。実際の長さはSamplerのChunksとSeconds per Chunkで決まります。
旧Workflowは開発履歴としてsource repositoryへ保持しますが、V3.8 Registry packageからは除外します。旧保存Workflowは[Migration Policy](V38_RELEASE_AND_MIGRATION.md)に従い、対応するhistorical Release/tagで開いてください。Registryの正式pack検証・公開はGitHub sourceの更新とは別工程です。
## V3.4互換
V3.4の実装moduleは内部継承とhistorical testのためsourceへ保持しています。現行の登録面は上記12 IDに限定します。それ以外のIDを必要とするV3.4保存Workflowは、対応するhistorical packageを使用してください。
## V3.4.0 Stable
V3.4.0は、最大3枚のReference Image、Run Storage、Spectrum Interop、Decode後のAudio / Video Seam補正を維持し、`Driving Audio`と`Video Guide Frames`を正式な入力として追加します。
`Driving Audio`は入力音声を絶対時間で各Chunkのガイドに使い、最終出力には元音声を維持します。`Video Guide Frames`はVideo loaderが出力したIMAGE frame batchを全Chunkで継続使用し、`Efficient - 0.4 MP`、`Balanced - 0.6 MP`、`Match Output`から処理解像度を選択できます。青いIMAGE socketですが、通常の1枚の静止画参照ではなく動画フレーム列を接続します。
### Size項目の条件付き表示と内部Resize
V3.5のCompact UIでは、対応する入力を接続した時だけSize項目を表示します。
- `reference_image_1`~`reference_image_3`のいずれかを接続すると、`Reference Size`を表示します。
- `Video Guide Frames`を接続すると、`Video Guide Size`を表示します。以前のWorkflowやIssueで`Video Reference Size`と記載されている項目と同じもので、V3.5.1では表示名だけを`Video Guide Size`へ明確化し、既存のbackend keyと保存済みWorkflowの互換性を維持しています。
したがって、新規ノードで見えないSize項目は欠落ではなく、未接続のため非表示になっています。Reference ImageとVideo Guide FramesはVAE Encode前に内部Resizeされます。Video Guide Framesはアスペクト比を維持してH3 canvasへ整列し、縮小が必要な場合はLanczosを使用します。小さい入力を自動的に拡大はしません。通常は外部Resizeノードを必要としませんが、意図的なcrop、小さい素材の強制upscale、独自のresize/upscale方式を使う場合には外部処理を利用できます。動画のdecodeとframe-rate変換は引き続きloader側の責任です。
Reference入力のバイパスされたソケットは無視され、有効な画像だけがPicture 1から連続採番されます。空欄や不完全なプロンプトも停止せず、警告とFixed fallbackで生成を続行します。未知のモデル、ノード、プロンプト形式を独自判断で拒否する制限も撤廃しました。
## Run Storage:第2部分だけを再生成
通常のT2VA/I2VAを`2 chunks x 5 seconds`にした場合、Chunk 1が最初の5秒、Chunk 2が次の5秒です。
初回生成前に次を設定します。
1. `Run Storage`を`Save + Auto Resume`にします。
2. 固定の`Run Name`を選び、Base Seedも固定します。
3. まず10秒全体を一度生成します。Continuumは完了した各Chunkのraw Video/Audio latentを保存します。
最初の5秒を保持して、第2部分だけを作り直す手順は次のとおりです。
1. Run Name、Base Seed、model、LoRA、解像度、sampler、SIGMAS/steps、Reference、Continuity設定を初回と同じに保ちます。
2. `Regenerate From`を`Chunk 2`にします。
3. `Variation Nonce = 0`なら、完了済みの再生成に対して新しいバリエーションを自動選択します。互換性のある未完了runを再開する場合は、既存のバリエーションを引き継ぎます。`1以上`は指定値を固定するため、同じ入力で別の結果を求める場合は別の正の値へ変更します。Base Seedは固定のままにします。
4. Workflowをもう一度Queueします。
Review画面の`Try this chunk again`は、この手動指定とは別に、内部のバリエーションを自動で進めます。
Continuumは保存済みChunk 1を再Samplingせずに読み込み、その確定済み末尾をContinuation contextとして新しいChunk 2をSamplingします。Reportの目安は`1 reused, 1 generated`です。その後、10秒全体を再Decode/Assemblyします。第1ChunkのSampling結果は再利用され、境界のSeam処理は新しいAssembly時に再評価されます。
第1部分と第2部分で異なる指示を使う場合は、初回生成前にListまたはTimeline Promptとして分けておきます。保存後にPrompt Planや生成Contractの入力を変更すると新しいRevisionになり、以前のChunk 1を再利用できない場合があります。初回を`Run Storage = Off`で生成した場合も、後から遡って再利用することはできません。`Regenerate From`はChunk境界単位であり、1つのChunk内部の任意領域を編集する機能ではありません。
FL2VA Long Terminal Mergeは重要な例外です。最後の2つのlogical Chunkを1つのatomicなphysical Samplingとして扱うため、その2つは一緒に再生成されます。
Run Storageの保存契約にはDriving AudioとVideo Guide Framesのidentityを含め、入力変更後に古いChunkを誤再利用しないようにしています。
## Driving Audio / Video Guide Frames / Seam
Driving Audioだけを使う、Video Guide Framesだけを使う、Video Guide Framesと別音声を組み合わせる、という3通りの接続に対応します。Video Guide Framesは24 fps入力を基準とし、異なるfpsの素材はLoad Video側で24 fpsへ設定してください。
H3 Continuum Assemble + SeamはAudio Seam AutoとVideo Seam Autoを既定とし、フレーム削除を行わずに境界の一時的な露出・色変化を補正します。Auto 2は露出ランプ向けの実験的な追加モードです。
## Legacy V2.1.7 Stable Facade UI
V2.1.7では、V2.1.6の動的表示制御を廃止し、標準ComfyUI描画だけを使用する静的Facadeへ移行しました。生成処理は互換用`H3ContinuumSamplerV2`を再利用し、State、Session、Prompt Plan、Spectrum Interop、保存Schemaは変更していません。
通常表示は次を中心にします。
```text
model / clip / video_vae / audio_vae / sampler / sigmas
Sequence Prompt
first_frame
prompt_overrides(任意pack)
advanced(任意pack)
Prompt Format
chunks
chunk_seconds
width / height
continuity
base_seed
control after generate
Seam Correction
Seam Correction
outputs: images / audio / result
```
### 補助ノード
- `H3 Continuum Clip Overrides`: Clip別Promptを1本のpackへまとめます。
- `H3 Continuum Advanced`: resume/session入力と低頻度設定をまとめます。
- `H3 Continuum Result`: resultから`last_state / session / report`を展開します。
- 補助ノードを使わない通常生成では、Sampler本体だけで動作します。
### Legacy互換
当時のV2.1.7では、旧`H3ContinuumSamplerV2`をLegacy/Coreノードとして登録し、既存WorkflowのノードID、入力順、出力順を維持していました。これはV3.8の公開範囲を示す説明ではありません。このIDを使う旧Workflowは対応するhistorical Release/tagで開いてください。
## Continuity
- `Auto — conservative`
- `Balanced — 22 frames`(標準)
- `Fast — 5 frames`
- `Strong — 39 frames (Experimental)`
5/22/39フレームはH3の時間latent周期`1,4,4,4,4`に対して、それぞれ2/7/12 temporal latent stepsです。
## Spectrum
Continuumは有効な前チャンクContextがある場合だけ`h3_continuum` Interop API v1をMODEL optionsへ付与し、対応Spectrumへ`min_actual_prefix_steps=2`を要求します。初回チャンクにはInterop keyを付与しません。未知・未対応Spectrumでは通常Spectrumとしてfail-openします。
## 推奨配線
```text
UNET -> SageAttention -> Sol-Attn -> LoRA -> Spectrum -> H3 Continuum Sampler
```
まず`chunks=2 / chunk_seconds=5 / Balanced 22f / Seam Correction=Off`でNative Continuityを確認し、その後`Seam Correction=Auto`を比較してください。
## インストール
V3.8はComfyUI 0.34.2で検証しています。以下の歴史的な受入記録には旧ComfyUI版も登場しますが、現在のインストール目標ではありません。V3.8の依存を満たす限り、ComfyUIの最新版は必須ではありません。
ZIPを展開して、`ComfyUI-H3-Continuum`フォルダーを`ComfyUI/custom_nodes/`へ置き、ComfyUIを再起動します。
再起動後はV3.8X2 Workflowで`H3 Continuum Sampler V3.8`、V3.9 Workflowで`H3 Continuum Sampler V3.9`と`H3 Continuum Reference Images V3.9`を確認してください。現行の登録面は計12 IDです。
旧`ComfyUI-H3-Continuum-Join`が残っている場合は同時ロードを避けるため削除または退避してください。同梱`install_windows.bat`は旧名・新名の既存フォルダーを日時付きでバックアップします。
## 検査
Public Surface suiteはV3.8X2の10 IDとV3.9を含む計12 ID、公式Workflowと各ZIP内JSONの同一性、外部依存、Registry除外、既存V3.8 widget/socket順、`Show Advanced Settings`/`Hide Advanced Settings`による表示切替を確認します。Registry配布対象のハッシュは`REGISTRY_MANIFEST.sha256`、source側の整合性は`MANIFEST.sha256`で管理します。
**過去のV3.8X2準備証拠:** Full CPU suiteは**1,367 passed / 1 skipped / 0 failed**です。従来の最終GPU Functional Gateでは`3 × 5秒`のReview Each Chunkを実行し、Q1~Q3はphysical groupを1つずつ生成、Q4は`3 reused / 0 generated`で3 groupすべてを再利用しました。Q3とQ4の復号後RGBおよびPCM SHA-256は一致しています。9枚Reference Imageの別Gateは下記に記録しています。これは実行、prefix再利用、AV再構築、テストした9枚経路の確認であり、画像・音声の総合的な主観品質評価ではありません。
### 9枚Reference Image入力Gate(0.3 MP参照画像)
テストしたNVIDIA GeForce RTX 5060 Ti 16 GB/RAM 64 GB環境では、Reference Image 1~9、`512 × 608`出力、`3 × 5秒` Full Run、Spectrum Off、LoRA Off、Sage Attention、Balanced 22-frame continuity、各約`0.30 MP`のテスト参照画像で完走しました。Sampling中のVRAM使用量は、利用可能な`16,311 MiB`のうち約`15.0~15.5 GiB`であり、16 GB GPUでは余裕の小さい高負荷構成です。
9枚bundleは、追加の外観・ポーズ情報が必要で、十分なVRAM余裕を確保できる環境で使用してください。このsingle-seed Gateは実行と基本的な視覚的連続性を確認したものです。9枚が3枚より常に高画質であること、高解像度参照・大きな出力・別model・同時GPU負荷でも動作することを保証するものではありません。
16 GB GPUで9枚を使う場合は、読み込む前に元画像のコピーを約0.30 MPへ縮小してください。内部のconditioning用Resizeは継続しますが、9枚の大きな元画像によるメモリ負荷を管理する代わりにはなりません。
検証結果はテストしたローカルsourceと環境に限られます。導入済み環境が最新版であること、すべてのmodel/wrapperが任意のGPUに収まること、別のpromptでも同じ画質になることを保証するものではありません。
```text
python -m compileall -q .
pytest -q
```
MIT License。
+37
View File
@@ -0,0 +1,37 @@
# Regeneration selections and automatic variations — 3.9.6
## Operation and compatibility
V3.8 and V3.9 retain explicit Regenerate From and Variation Nonce across workflow save/reload and Queue submission. The intent marker is stored in workflow properties; public inputs, widget order and node IDs are unchanged. Choosing Auto clears the marker. The existing one-shot restart remains one-shot.
With progress saving enabled, choose Full Run and an explicit Chunk. Nonce 0 lets Run Storage select an automatic variation: an interrupted compatible variation may continue, and queuing again after completion creates another variation. Fixed positive nonces and ordinary Full Run retain their existing deterministic identity and Core caching.
Restart ComfyUI and refresh the browser after updating. Older unmarked Chunk 1 / nonce 0 values cannot be distinguished from old restart residue; legacy cleanup still restores Auto. To retain a deliberate Chunk 1 / nonce 0, select it again in the updated UI and save once. Old unmarked later-chunk or fixed-positive-nonce selections are retained. Do not delete Runs/Takes to update.
Sampling v6, Graph v4, Run Storage v3 and State/Session formats are unchanged. This patch does not force resume_safe, relax reuse checks, rewrite unsafe/v5 history or migrate workflows. Run Storage decides whether a prefix is compatible. Earlier accepted chunks and previous Takes remain protected. All 25 distributed JSON/ZIP assets retain their bytes.
## Verified scope
Full real-Core CPU gate: 1,908 passed / 2 skipped / 2 existing Inspector expected failures / zero unexpected failures; CUDA initialization attempts 0. Production frontend fixtures cover 84 scenarios including 22 new restoration cases. Real-Core CPU cache checks cover ordinary Full Run, automatic nonce 0 and fixed positive nonces.
Actual normal ComfyUI (Core 0.40.0, frontend 1.55.17): save a separate test workflow, reload the browser, allow real Run history to load, and verify Chunk 2 / nonce 5 and Chunk 1 / nonce 0 remain selected. The saved JSON includes the matching intent marker. This browser test is distinct from actual GPU prompts submitted through the normal Core API.
| New-Run GPU Queue | Result |
| --- | --- |
| Review / Auto / nonce 0 | Generate Chunk 1 |
| Separate Queue, same Review input | Reuse Chunk 1, generate Chunk 2 |
| Full Run / Chunk 2 / nonce 5 | Keep Chunk 1, regenerate Chunk 2 |
| Full Run / Chunk 2 / nonce 0 | Keep Chunk 1, create automatic variation 6 |
| Identical Full Run / Chunk 2 / nonce 0 again | Keep Chunk 1, create automatic variation 7 |
GPU: RTX 5060 Ti, PyTorch attention, Spectrum OFF, 512x512, Euler/simple, 8-step Turbo, two 5-second chunks, no First/Last/Reference inputs. Earlier Chunk 1 SHA-256 stayed identical; every newly generated Chunk 2 differed. The last two prompts, including output prefix, were exactly identical and the Sampler was absent from execution_cached. All manifests recorded Sampling v6 and resume_safe=true. All five video/audio outputs (first: 5 seconds; others: 10 seconds) fully decoded at 24fps.
Existing saved workflows and reachable archived Run/Take files retained their recorded hashes. Older output paths unavailable before this repair were not counted as reverified; no missing-data diagnosis is inferred. Other attention/Spectrum combinations were not newly GPU-tested for this patch. Subjective audiovisual quality, mobile performance, full default-workflow acceptance and the known Inspector reconfigure issue remain separate. Previous Issue #31 coverage: ISSUE31_SAMPLING_SIGNATURES.md.
## 日本語
明示したRegenerate FromとVariation Nonceを保存・再読込後も保持します。保存ON/Full Run/明示Chunk/Nonce 0では、毎QueueでRun Storageが継続か新Variationかを判断します。完了後の同一入力でも再生成し、前のChunkと既存Takeは保持します。固定の正Nonceと通常Full Runのキャッシュは維持します。
ComfyUI再起動とブラウザー再読込が必要です。旧markerなしChunk 1/Nonce 0は、更新後にChunk 1を選び直して一度保存してください。旧unsafe履歴は自動修復しません。既存WorkflowやRun/Takeの削除は不要です。
実画面でChunk 2/5とChunk 1/0の保存再読込保持、実GPUでChunk 1→2継続・Chunk 2再生成・同一Full Run入力2回の自動Nonce 6→7、元Chunk 1保持、動画5本全デコードを確認しました。今回のGPU範囲はPyTorch/Spectrum OFFです。Sampling v6の安全判定、既存配布Workflow、旧Takeを保護しています。
+67
View File
@@ -0,0 +1,67 @@
# 3.9.10 — Spectrum reuse and Reference assignment display
Fix two causes of false **Sampling v6 read-only** in the supported Spectrum
configurations: Windows CRLF source line endings and Python 3.12 implementation
signatures. Also correct the Reference Images assignment table when Chunks comes
from an external INT node.
## Changes
- Audit Spectrum CRLF and LF source line endings identically, including module
closures. All other source bytes and the exact supported file sets stay checked.
- Add separate audited Python 3.12 profiles for upstream Spectrum 0.2.29 at
`806fe498d574a726fecb8818917f00a4b993fd77` and the existing audited local
implementation. Retain the original Python 3.13 profiles byte-for-byte.
- With external Chunks inputs, display **assignment range 1–16** instead of a
stale Sampler widget count. All/Off actions identify their range. This range
does not set generation length or automatically expand saved selections.
- Preserve public node IDs, input/widget order, Sampling v6, Graph v4, existing
Takes and official workflow files. No chunk-LoRA or Project Manager feature is
included.
## Update
Update Continuum, restart ComfyUI and refresh the browser. Existing workflows need
no rewiring. Keep saved Runs/Takes. For a newly supported configuration, start a
**new Run**: older unsafe/v5 Takes remain readable and are not automatically made
reusable. Normal supported Python 3.13 generation identities stay unchanged.
## Verification and limits
- Full local CPU/JS gate: **2,096 passed / 2 skipped / 2 known Inspector expected
failures**; compilation and offline registration of all twelve nodes passed.
- Targeted Spectrum/Attention: **169 passed / 1 skip**. Python 3.12 component:
**112 passed / 1 skip**, with 21 independent older-Core/Attention cases excluded.
- Separate-process storage: **24 cases**, comprising 16 PyTorch/Kitchen cases on
Python 3.13 and eight Spectrum/default-model cases on Python 3.12. All matched
the actual Take read-only JavaScript. Original raw data was retained, and all
six previously valid full Sampling Contracts stayed identical.
- Reference table browser checks passed external INT and KJ Set/Get routing,
selection controls, disconnect/restore, save/reload and workflow tab switching.
- **No new GPU acceptance is claimed for 3.9.10.** These are CPU and UI checks,
not Attention-kernel, H3 inference, image-quality or performance measurements.
Unknown source, changed functions, foreign closures and additional runtime state
remain non-reusable. The repair does not guarantee every Core/Kitchen/KJ/Sage or
Python combination. An independent Attention-observation failure in an older
Core/Kitchen environment can still disable reuse even when Spectrum is recognized.
The reporting user's exact environment has not been reproduced.
See the [compatibility guide](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/blob/v3.9.10/docs/ISSUE31_SAMPLING_SIGNATURES.md) for audit details.
## 日本語
対応済みSpectrum構成で、WindowsのCRLF改行とPython 3.12の実装識別によって
新規Takeまで「Sampling v6 read-only」になる判定を修正しました。外部INTでChunksを
指定したときのReference Images割り当て表も修正し、範囲1–16で設定できるようにしました。
生成チャンク数や既存の選択内容を自動変更する機能ではありません。
更新後はComfyUIを再起動し、ブラウザーを更新してください。Workflowの配線変更や
履歴削除は不要です。新しく対応した構成は新規Runで確認してください。
既存3.13の正常な識別値と、旧unsafe/v5 Takeの閲覧を維持します。
CPU/JS全回帰2,096成功、Python 3.12対象112成功、独立した保存・継続24ケースと
実JS照合、Reference表示のブラウザー操作を確認しました。
**今回の修正としてのGPU受入は未実施です。** 未知の実装や追加状態の拒否は維持し、
すべてのCore/Kitchen/KJ/Sage構成への対応や、報告者本人の環境再現を保証しません。
4.0はメジャーアップデート用として温存し、SemVer互換の3.9.10で公開します。
+21
View File
@@ -0,0 +1,21 @@
# v3.9.5 — Issue #31 Attention and Spectrum Take reuse
Fix finite identity observation of Core Model Attention Backend and slots-based SpectrumH3Binding so compatible new Runs can continue Review Each Chunk. Core PyTorch and Comfy Kitchen INT8 overrides include loaded implementations, closures and both container paths. Spectrum includes every effective config value and an explicit loaded-implementation contract; transient history/counters do not change identity.
Includes a dedicated profile for upstream Spectrum v0.2.29 at 806fe498d574a726fecb8818917f00a4b993fd77, plus the prior audited implementation. Other source/runtime combinations remain non-reusable when their identity is unknown. Restart ComfyUI and start a new Run after updating; previously unsafe/v5 history is preserved and is not promoted or repaired.
Validation: full local CPU gate 1,893 passed / 2 skipped / 2 existing Inspector expected failures / 0 failures; CUDA initialization attempts 0. Targeted signature gate: 152 passed / 1 scope-specific skip. Upstream 0.2.29 PyTorch+Spectrum and Kitchen+Spectrum both persisted Chunk 1 and reused it from a separate Queue to generate Chunk 2, with matching revision/contract, unchanged prefix hashes and fully decoded 512x512, 24fps, 10-second audio/video output. Actual CUDA backend calls were observed in the isolated gate. Prior normal-user-runtime checks also passed for the original audited Spectrum implementation.
All 25 existing JSON/ZIP assets and 259 pre-existing Run/Take files were preserved. Sampling v6, Graph v4, Run Storage v3, twelve public IDs and official workflow bytes remain unchanged. This is targeted functional acceptance; browser interaction, subjective audiovisual continuity, BSA/Flow/Untwist, KJ/Sage patch-node combinations and benchmarks remain separate. See docs/ISSUE31_SAMPLING_SIGNATURES.md for environment-specific support.
GitHub and Registry archives are checked for exact file sets, CRC and complete Manifest hashes. Hosted main and tag tests must pass before publication. Registry scan approval and real Manager installation remain separate from upload/archive verification. Prior tags, Releases and assets are retained.
## 日本語
Model Attention BackendとSpectrumの署名を正確に識別し、新しいRunのReview Each Chunk継続を修正します。公式Spectrum v0.2.29も専用の監査済みプロファイルで対応します。更新・再起動後は新しいRunを作成してください。以前のunsafe/v5履歴は保持し、自動修復・再利用可能への昇格は行いません。
全CPU回帰は1,893成功・2スキップ・既知Inspector問題の期待失敗2件、失敗0・CUDA初期化0。公開Spectrum v0.2.29のPyTorch/Comfy Kitchen両条件で、別QueueからChunk 1を再利用してChunk 2へ継続しました。署名・元Takeのハッシュ保持・完成10秒動画の全デコードと実CUDA呼出しを確認しています。
既存25資産と259件のRun/Takeファイル、公開12ノード、Workflow、Sampling v6を保護しました。未知のSpectrumや異なる実行環境は安全側で再利用を拒否します。実画面・目視画質/音声・BSAやKJの組合せ全般の合格を主張するものではありません。
GitHub本体ZIP・チェックサム、Registry配布ZIPの整合性は公開工程で別々に確認します。Registry側の検査承認と実Manager導入は別判定です。3.9.4/3.9.3/3.9.2などの元タグ・Release・配布物は保持します。
+21
View File
@@ -0,0 +1,21 @@
# v3.9.6 — Regeneration intent and automatic variations
Fix Regenerate From / Variation Nonce restoration in V3.8 and V3.9. Explicit Full Run regeneration with progress saving and nonce 0 now reaches Run Storage on every Queue, so repeating the same input after a completed variation creates another Take while preserving earlier accepted chunks. Ordinary Full Run and fixed positive nonces keep their caching behavior.
Restart ComfyUI and refresh the browser after updating. For an older unmarked Chunk 1 / nonce 0, select Chunk 1 again and save once. Existing Runs/Takes and official workflows do not need deletion or migration. Sampling v6 safety and unsafe/v5 history guards are unchanged.
Validation: 1,908 local CPU passed / 2 skipped / 2 known Inspector expected failures / zero unexpected failures and CUDA initialization attempts. Actual browser save/reload retained Chunk 2 / nonce 5 and Chunk 1 / nonce 0 after history loading. Five new-Run GPU queues passed Chunk 1 creation, separate-Queue Chunk 2 continuation, Chunk 2 regeneration and two identical Full Run / nonce 0 submissions (automatic variations 6 then 7). Earlier Chunk 1 bytes were unchanged; all five video/audio outputs fully decoded.
GPU scope: PyTorch attention, Spectrum OFF, RTX 5060 Ti, 512x512, 8-step Turbo, two 5-second chunks. Browser acceptance and Core-API GPU submissions are distinct checks. Subjective quality, other accelerators and the known Inspector issue are not newly accepted. Existing saved workflows, reachable archived Takes and all 25 JSON/ZIP distribution assets were hash-protected. Previously unavailable old output paths are not claimed as reverified. Details: docs/REGENERATE_INTENT_V396.md.
GitHub ZIP/checksums and Registry archive pass exact file-set, CRC and Manifest checks during publication. Registry upload, service scan approval and actual Manager installation are separate. Earlier Releases/tags/assets are retained.
## 日本語
V3.8/V3.9の再生成指定・Nonceを保存再読込後も保持します。保存ON/Full Run/明示Chunk/Nonce 0の同一入力でもRun Storageへ到達し、完了後は新Variationを生成できます。前のChunkと既存Takeを保持し、通常Full Runと固定Nonceのキャッシュは維持します。
更新後はComfyUIを再起動し、ブラウザーを再読込してください。旧markerなしChunk 1/Nonce 0は選び直して一度保存してください。既存WorkflowやRun/Takeの削除・移行は不要です。Sampling v6の安全判定と旧unsafe履歴の扱いは変更しません。
CPU 1,908成功・2スキップ・既知Inspector期待失敗2、失敗0・CUDA初期化0。実画面でChunk 2/Nonce 5とChunk 1/Nonce 0の保存再読込保持を確認。PyTorch/Spectrum OFFの新RunでGPU 5 Queueが成功し、Chunk 1→2継続・Chunk 2再生成・同一入力の自動Nonce 6→7・元Chunk 1保持・動画5本全デコードを確認しました。画質評価や他アクセラレーターの新規合格は主張しません。
GitHub配布物整合性とRegistryアップロードは公開工程で確認します。Registry検査承認・実Manager導入は別判定です。以前の元タグ・Release・配布物を保持します。
+30
View File
@@ -0,0 +1,30 @@
# v3.9.7 — External-input regeneration cache fix
Fix automatic Full Run variations when Chunks or Variation Nonce come from external nodes. Unresolved linked inputs no longer let Core skip an explicitly requested regeneration. Earlier accepted chunks, existing Takes, public inputs and official workflows are preserved. Update the custom nodes, restart ComfyUI and refresh the browser; no workflow changes or history deletion are required.
CPU: 1,948 passed / 2 skipped / 2 known Inspector expected failures; zero unexpected failures and zero CUDA initialization attempts. Targeted Core-API GPU queues verified external Chunks=2 / Nonce=0, repeated identical automatic variations 1 then 2, and 1 reused / 1 generated chunk per variation. Browser save/reload and Run-button submissions were checked separately. GPU scope: PyTorch attention, Spectrum OFF, 512x512, 8-step Turbo, two 5-second chunks. This is a targeted cache repair, not general image-quality or all-accelerator acceptance.
Known ordinary Full Run, storage OFF and directly specified fixed nonces retain stable caching. An unresolved externally connected nonce may reevaluate the Sampler even when fixed; Run Storage still decides safe prefix reuse. Sampling v6 / Graph v4 and existing unsafe/v5 history safeguards are unchanged.
Existing workflow JSON/ZIP assets remain byte-for-byte unchanged. GitHub asset integrity and Registry upload are verified by the publishing workflow; Registry scan approval and Manager installation remain separate checks.
## 日本語
ChunksやVariation Nonceを外部ノードから接続した際に、明示したFull Run再生成がCoreのキャッシュで省略される問題を修正しました。前の採用済みChunk、既存Take、公開入力、公式Workflowは維持します。更新後はComfyUIを再起動し、ブラウザーを再読込してください。Workflowの変更や履歴の削除は不要です。
CPU 1,948成功、2スキップ、既知の期待失敗2件。外部INT接続で同一入力から自動Variationが進み、Chunk 1を保持してChunk 2だけ再生成することをGPUで確認しました。ブラウザーの保存・再読込とRun操作も別途確認しています。画質・全アクセラレーター構成の新規合格は主張しません。外部接続の固定Nonceはキャッシュ判定時に値が未解決のためSamplerを再評価する場合がありますが、保存済みprefixの再利用は既存の安全判定に従います。
## Post-release verification on the published v3.9.7 tag
**Published v3.9.7 GPU recheck (2026-10-15): PASS.** After restarting with the published tag (`18df80d`), external Core INT nodes supplied Chunks=2 and Nonce=0. Two completely identical Chunk 2 regeneration prompts advanced automatic Nonce **1 → 2**; both reused Chunk 1 and generated only Chunk 2. Earlier Takes and raw latents were retained. All three outputs (baseline plus two variations) fully decoded with 240 frames and 10-second video/audio. Existing 60 workflows, one settings file, 30 Run/Take files and all 459 release files remained hash-identical. Scope: RTX 5060 Ti 16 GB, 512×512, 24 fps, 2×5 seconds, 8-step Turbo, PyTorch attention, Spectrum OFF, Core API submissions. No CUDA/OOM/execution errors were observed; this does not extend browser, subjective-quality or other-accelerator acceptance.
| Queue | Reused chunks | Generated chunks | Automatic Nonce | Elapsed |
| --- | ---: | ---: | ---: | ---: |
| Baseline | 0 | 2 | 0 | 177.8 s |
| Regenerate Chunk 2, first run | 1 | 1 | 1 | 90.4 s |
| Same input, second run | 1 | 1 | 2 | 90.5 s |
### 公開版での追加GPU検証
**公開版3.9.7でのGPU再検証(2026-10-15):PASS。** 公開タグ(`18df80d`)で再起動後、外部Core INTからChunks=2・Nonce=0を入力し、完全に同一のプロンプトでChunk 2を2回再生成しました。自動Nonceは**1→2**へ進み、両回ともChunk 1を再利用してChunk 2だけを生成。以前のTake・raw latentも保持しました。基準生成を含む動画3本は240フレーム・映像/音声各10秒で全デコード成功。既存Workflow 60件・設定1件・Run/Take 30ファイル・公開版459ファイルのハッシュは不変です。条件はRTX 5060 Ti 16 GB、512×512、24fps、2×5秒、8-step Turbo、PyTorch attention/Spectrum OFF、Core API経由。CUDA/OOM/実行エラーは確認されず、ブラウザー・主観画質・他アクセラレーターの受入範囲は拡張していません。
+47
View File
@@ -0,0 +1,47 @@
# v3.9.8 — KJ Model Preview Override continuation
Fix saved-Take continuation after an explicitly audited KJ Preview Override retains preview images. The dedicated adapter separates reviewed transient frames from fixed settings and implementation identity, preserving cross-key/list wrapper execution order. A separate Queue can reuse Chunk 1 and generate only Chunk 2. Sampling v6 safety, existing Takes, public inputs and official workflow bytes are preserved.
**Limited Phase A support:** Preview Override VAE / Audio VAE disconnected, Tiny VAE `none`, Core NoPreviews/Auto/Latent2RGB, and only the exact provider/Core audit profile. KJNodes 1.5.2 alone is not sufficient; the audited local Preview Override source differs from upstream. Unknown or altered providers retain conservative non-reuse. See [the supported source hashes and update guide](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/blob/v3.9.8/docs/KJ_PREVIEW_OVERRIDE_V398.md).
Restart ComfyUI, refresh the browser and start a new Run. Do not delete old workflows or Takes. Earlier safe KJ identities may differ; unsafe/v5 history is not repaired or promoted. Non-KJ v6 signatures are unchanged.
CPU: 1,995 passed / 3 skipped / 2 existing Inspector expected failures; zero unexpected failures or CUDA initialization attempts. Dedicated real-provider tests: 48 passed, no skips. Four normal-runtime Core-API GPU Queues passed: PyTorch and Comfy Kitchen INT8 + audited local Spectrum, each saving Chunk 1 and continuing to Chunk 2 in a separate Queue with held previews. The KJ MODEL was cached on each second Queue, contracts stayed equal, and original Chunk 1 bytes were preserved. All four AV outputs fully decoded.
GPU scope: RTX 5060 Ti 16 GB, 512x512, 24fps, 8-step Turbo, 2x5-second chunks, Core Auto preview, no preview VAE. The initial GPU gate used candidate product bytes; the subsequent published-tag recheck is recorded below. No new browser Run-button, per-kernel return trace, subjective-quality or performance acceptance is claimed. The Inspector issue remains. Source/tag CI and exact GitHub/Registry archive checks run in publication; Registry scan approval and real Manager installation are separate.
## 日本語
監査済みKJ Model Preview Overrideで画像を保持した後、保存Takeを継続できない問題を局所修正しました。一時画像と固定設定を分離し、実装識別とラッパー順序、Sampling v6の安全判定を維持します。
対応は固定した監査プロファイルと、VAE/Audio VAE未接続・Tiny VAE none・Core NoPreviews/Auto/Latent2RGBに限定します。KJNodes 1.5.2全般を保証するものではなく、監査したローカルソースは公開upstreamと異なります。未対応構成は安全側で再利用不可を維持します。
更新後はComfyUIを再起動し、ブラウザーを再読込して新Runを作成してください。既存Take・Workflowは保持し、旧unsafe履歴を自動修復しません。CPU 1,995成功(専用48件を含む)、PyTorchおよびKitchen+ローカルSpectrumの4 Queueで別QueueのChunk 1→2継続、元Chunk 1保持、動画4本の全デコードを確認。画質・速度・全KJ構成の合格は主張しません。Registryの検査承認と実Manager導入は別確認です。
## Post-release verification on the published v3.9.8 tag
### Published v3.9.8 GPU recheck — 2026-10-16: PASS
After restarting the normal ComfyUI installation with the published tag **`v3.9.8` / `58fbd855`**, all 464 installed release files matched the tag. The startup log confirmed 3.9.8, and new saved Sampling v6 contracts contained the KJ adapter and pinned audit-profile digest. Two brand-new Runs passed four Core-API GPU Queues:
| KJ Preview Override configuration | Separate-Queue Chunk 1 → 2 |
| --- | --- |
| Model Attention Backend: PyTorch, Spectrum OFF | **PASS — 1 reused / 1 generated** |
| Model Attention Backend: Comfy Kitchen INT8 + audited local Spectrum | **PASS — 1 reused / 1 generated** |
On both second Queues, the KJ MODEL stayed cached with preview images retained (**8 → 16**), while the Sampler executed. The contract/revision and original Chunk 1 latent hashes stayed unchanged; `resume_safe=true`. All four video/audio outputs fully decoded (5 seconds / 120 frames, then 10 seconds / 240 frames, 512×512 at 24fps). Existing workflows and Takes were preserved: 170 protected files, including 71 saved workflow files and 46 pre-existing Run/Take files, retained their hashes.
**Scope:** only the exact audited KJ/Core profile; the local KJ source differs from upstream, so KJNodes 1.5.2 alone is insufficient. Preview Override VAE / Audio VAE were disconnected, Tiny VAE was `none`, Core preview was Auto. GPU: RTX 5060 Ti 16 GB, 8-step Turbo, Euler/simple, 2×5-second chunks. Spectrum used the audited local offline-replay implementation, system-RAM history/archive and model-aware OFF. **Unverified:** browser Run-button/save-reload behavior, the reporter's exact complete graph, other KJ/Spectrum builds, SolAttn/preview VAE paths, per-kernel return tracing, subjective quality and speed/VRAM benchmarks. Retained PNGs were 32×32, not high-resolution preview-quality evidence. The known Inspector issue remains; old unsafe/v5 Takes are not repaired or promoted. [Issue #31 result](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/issues/31#issuecomment-6209571261).
### 公開版v3.9.8のGPU再検証 — 2026-10-16:PASS
通常ComfyUIに公開タグ **`v3.9.8`/`58fbd855`** を反映して再起動し、公開版464ファイルとの一致、実ログの3.9.8読込、新しいSampling v6契約内のKJアダプターと固定監査プロファイルを確認しました。新Runを2つ使用した計4 Queueの結果です。
| KJ Preview Override併用構成 | 別QueueのChunk 1→2継続 |
| --- | --- |
| Model Attention Backend:PyTorch/Spectrum OFF | **PASS — 1再利用/1生成** |
| Model Attention Backend:Comfy Kitchen INT8+監査済みローカルSpectrum | **PASS — 1再利用/1生成** |
両構成の2回目は画像を保持したKJ MODELをキャッシュ再利用(**8→16枚**)し、Samplerは実行。契約・revision・元Chunk 1 latentのSHA-256を維持し、`resume_safe=true`でした。映像・音声4本は全デコード成功(512×512/24fps、5秒120フレーム→10秒240フレーム)。保護対象170ファイル(保存Workflow 71件・既存Run/Take 46ファイルを含む)は全ハッシュ不変です。
**対応範囲:** 固定したKJ/Core監査プロファイルのみ。ローカルKJソースはupstreamと異なり、KJNodes 1.5.2全般の合格ではありません。Preview OverrideのVAE/Audio VAE未接続・Tiny VAE `none`・Core Auto、RTX 5060 Ti 16 GB、8-step Turbo/Euler/simple、2×5秒。Spectrumは監査済みローカル版のoffline replay、履歴・archiveはsystem RAM、model-aware OFFです。**未確認:** ブラウザーRunボタン/保存再読込、報告者の完全一致graph、別KJ/Spectrum版、SolAttn/preview VAE、カーネル単位の戻り値、主観画質・音質、速度・VRAM比較。保持PNGは32×32で、高解像度プレビュー品質の証明ではありません。既知Inspector問題は残り、旧unsafe/v5 Takeを自動修復・昇格しません。[Issue #31への結果返信](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/issues/31#issuecomment-6209571261)。
+69
View File
@@ -0,0 +1,69 @@
# 3.9.9 — Public KJ Preview Override and Review continuation
Some new Runs were marked **Sampling v6 read-only** when using the upstream KJ
Model Preview Override implementation, even though generation completed. This
release adds a separately audited upstream profile and explains saved reuse
failures in Review and Render History. The original supported local profile and
unaffected generation identities are retained; the safety checks are not relaxed.
## Update
Update Continuum, restart ComfyUI and hard-refresh the browser. Existing workflows
need no rewiring. Keep all saved Runs/Takes. After correcting an unsupported
configuration, start a **new Run**: old unsafe/v5 Takes remain readable and are
not automatically promoted to reusable history.
Supported upstream KJ commit: `17195fcfa8a05f601b766a0f09c9733284f47416`.
On **Model Preview Override**, leave its optional VAE and Audio VAE disconnected,
set Tiny VAE to `none`, and use Core NoPreviews, Auto or Latent2RGB. Keep the
normal generation VAEs connected. This is not blanket support for every KJ/Core
version or unknown wrapper. See [the exact profiles and recovery steps](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/blob/v3.9.9/docs/KJ_PREVIEW_OVERRIDE_V398.md).
## Validation
- CPU: **2,075 passed / 2 skipped / 2 known Inspector expected failures**, zero
unexpected failures. The 90 production-JS cases and eight actual Run Storage
vectors passed; these overlap the pytest total. Offline twelve-node,
PackedLayout and Fixed 3x5s checks passed with zero recorded CPU CUDA
initialization attempts.
- Upstream KJ GPU: two separate Queues, **512x512 / 24fps / 2 x 5 seconds**,
FL2VA INT8 convrot, Qwen3VL NVFP4 AWQ, 8-step Turbo/Euler/simple and explicit
PyTorch attention. Chunk 1 paused for Review; the next Queue reused it and
generated only Chunk 2. The same cached KJ wrapper held eight real preview
images before Continue, while MODEL and sampling signatures remained equal.
- Browser on normal ComfyUI: with the pre-existing audited local KJ profile,
restart and frontend reload succeeded. **Use This Take** and **Continue From
Here** were visibly enabled. Continue From Here followed by the browser Run
button completed Chunk 2: **one reused / one generated**. Core execution times
were approximately 100 and 92 seconds, including decode/save; they are not a
speed comparison. Opening Render History involved one manual click; the
subsequent Continue selection and Run submission were automated browser actions.
- Both GPU pairs preserved the original Chunk 1 raw SHA-256, original Take and
correct parent link. All four 5s/10s video/audio outputs fully decoded. The
upstream GPU pair also passed saved-latent finite-value checks. No execution
errors or OOM occurred. The observed upstream system VRAM maximum was 15,446
MiB at roughly six-second intervals, not an instantaneous per-process peak.
- All 21 existing workflow JSON/ZIP files remain byte-identical. Existing user
Take/workflow files were preserved; there is no storage or schema migration.
These are targeted continuation/Review checks on Core 0.42.1, frontend 1.55.18,
Python 3.13.12, torch 2.14.1+cu130 and RTX 5060 Ti 16GB. Full CPU validation also
uses real Core/provider imports. They do not establish visual/audio quality,
long-video quality, all accelerators, all KJ versions, or save/reload acceptance
of every shipped workflow. The separate Turbo Sample already present on main is
included unchanged; its six-step/4-step-LoRA pairing has no new GPU acceptance.
## 日本語
公開KJ Preview Overrideを接続した新規Runまで「v6 read-only」になる対応漏れを修正し、
再利用不可のTakeには保存時の理由と対処法を表示します。安全判定は緩めていません。
更新後はComfyUIを再起動し、ブラウザーを強制再読込してください。既存Workflowの配線変更や
Run/Take削除は不要です。旧unsafe/v5 Takeは保持されますが、自動的に再利用可能には
なりません。設定修正後は新Runで開始してください。Preview Override側の任意VAE/Audio VAEは
未接続、Tiny VAEはnoneが条件です。通常の生成用VAEは接続したまま使用します。
CPU 2,075成功、公開KJでの実GPU継続、既設KJを維持した実ブラウザーでのボタン有効化と
Continue From Here→Run→Chunk 2完了を確認しました。両GPU構成でChunk 1を保持し、
1再利用・1新規生成、Takeの親子関係、映像・音声全デコードを確認しています。
対象修正の検証結果であり、全KJ構成や画質・長時間品質への包括的な合格ではありません。
+44
View File
@@ -0,0 +1,44 @@
# V3.9 Turbo Sample
A separate, user-supplied example of Continuum prompt formatting, not a new model
or a verified quality preset. Available in Templates as
`MiniMax_H3_Continuum_V39_Turbo_Sample` (no parenthesized suffix).
- Prompt Format: Auto; 2 chunks x 8 seconds; Full Run; Progress OFF.
- The multiline prompt demonstrates `[Chunk 1]`, `[Chunk 2]`, first-image alignment,
continuous dancing with a fixed background/camera, and an off-screen male Thai rap.
`<d>[Thai] ... </d>` is H3 lyric syntax, not part of the template name.
- Saved model: `minimax/DasiwaMinimaxH3_dasiwaHybridV4_int8.safetensors`.
Obtain it from the [author's model page](https://civitai.com/models/2877206?modelVersionId=3391862)
or select a compatible installed H3 model. No model weights are bundled.
- One LoRA is enabled: `minimax/minimax_h3_fl2v_turbo_4step_v1.2_768p_comfyui_bf16.safetensors`,
strength 1, from [LightX2V](https://huggingface.co/lightx2v/Minimax-h3-Turbo).
The supplied graph uses Euler / simple / 6 steps / Draft 0.30 MP. These are
preserved sample values, not a validated 4-step/768p configuration; check the
selected model and LoRA instructions before generation.
- Video VAE: `minimax_h3_video_vae_int8_convrot.safetensors`; audio VAE:
`minimax_h3_audio_vae_fp32.safetensors`; text encoder:
`qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors`, type `minimax`.
See [model download links and folders](../README.md#required-models-and-extensions).
- Select your own First Image (`example.png` is only a placeholder). The disabled
optional audio/video inputs retain sample filenames; select real files before
enabling them. Reselect model/LoRA paths if you do not use a `minimax` subfolder.
- Requires KJNodes (SageAttention and Model Preview Override), rgthree, and Spectrum
node definitions. Spectrum is OFF. Preview Override has video/audio VAE connected;
this is outside the audited v3.9.8 Preview Phase A Take-reuse profile.
- Thai lyric accuracy, unchanged background, identity preservation and subjective
video/audio quality are not established by adding this sample. Thai is outside
H3's explicitly listed stable dialogue languages. No new GPU test is claimed.
- Existing standard workflows and saved Runs/Takes are unchanged. Outputs use the
separate `video/MiniMax_H3_Continuum_V39_Turbo_Sample` prefix. GitHub main includes
this sample; the already-published v3.9.8 release/Registry package is not replaced.
## 日本語
標準Workflowとは別のTurboサンプルです。Auto・2チャンク×8秒で、改行した
映像・環境音・音楽の記述、入力画像の指定、固定背景、画面外の男性タイ語ラップを
示します。画像・モデル・LoRAは各自で導入して選択してください。
4-step/768p LoRAに対して6 steps・Draft解像度という添付の設定を保持しています。
推奨設定や画質の合格を意味しません。タイ語の発音、人物・背景保持も未保証です。
KJ Preview OverrideのVAE接続はv3.9.8の限定Take再利用検証の範囲外です。
今回はサンプル追加とCPU検証のみで、実環境反映・新規GPU試験・Release更新は行いません。
+705
View File
@@ -0,0 +1,705 @@
# H3 Continuum — illustrated user guide
[Back to the quick start](../README.md) · [V3.9 References](V39_REFERENCE_INLINE.md) · [History](README_HISTORY.md)
This guide preserves the detailed operating instructions and screenshots from
the README at source commit `894869d9` (package 3.9.8).
V3.8-labeled controls and screenshots describe the retained V3.8X2 workflow;
the V3.9 setup and Reference sections identify the separate current workflow.
Earlier dated fixes, version statements and measurements are historical evidence,
not a new acceptance claim. Use the [current README](../README.md) and the
[v3.9.8 compatibility profile](KJ_PREVIEW_OVERRIDE_V398.md) for current scope.
- [V3.9 template setup](#start-here-v39-and-the-retained-v38x2-workflow)
- [Reference Images](#v39-reference-images)
- [Complete V3.8 control reference](#complete-v38-ui-reference)
- [Review, retry, Takes and extension](#review-continue-and-revise-a-chunked-video)
<!-- Preserved README sections begin. -->
## Start here: V3.9 and the retained V3.8X2 workflow
The ComfyUI **Templates → ComfyUI-H3-Continuum** gallery offers two separate entries: the user-selected [V3.9 workflow](../examples/workflows/MiniMax_H3_Continuum_V39.json) and [V3.8X2 workflow](../examples/workflows/MiniMax_H3_Continuum_V38X2.json). The [V3.8X2 workflow ZIP](../examples/workflows/MiniMax_H3_Continuum_V38X2.zip) contains **both unchanged JSON workflows** for one-download setup; the [V3.9 ZIP](../examples/workflows/MiniMax_H3_Continuum_V39.zip) contains only V3.9. These are workflow archives, not custom-node installers. The V3.8 Sampler remains available, so existing V3.8X2 workflows still load. V3.9 does not automatically convert their Reference connections or saved Runs/Takes: preserve the old workflow and open the dedicated V3.9 graph for new work.
The selected V3.9 template has 32 nodes. Image loaders 1–6 are wired to the new helper and bypassed by default; slots 7–9 remain available but require the user to add and connect their own image loaders. Choose actual files and enable only the desired images. The saved size controls are **First Image + Draft — 0.30 MP**, and the separate First Image loader is enabled; select your own First Image file before Queue. The saved 480×640 Width/Height values apply only in Manual mode. `Reference Image Size` is `Match Output`; it resizes conditioning after source-image load, not the source files themselves. For a clear per-chunk Reference test, turn First Image and Last Image off, switch to Manual, and use Reference Images alone. First/Last additionally constrain the video; Last Image may merge the final two logical chunks into one physical group, which requires the same effective Reference set in both. A subject may remain visible in later chunks through video continuity even after its Reference is no longer assigned.
In the helper, `All chunks` uses connected images throughout; `Per chunk` provides an image-by-chunk table. The template starts in `Per chunk` with all assignments off: after enabling a loader, tick the chunks that should use it. Slot numbers never shift: Image 1 is `@R1` and Image 9 is `@R9`, even when intervening slots are empty. In Sequence Prompt, describe the action for each chunk and use `@R` tags only for active References; their actual `<Picture N>` numbers are resolved per group. Unused tags warn without stopping generation. Start with short chunks and pre-resized images: source images may consume RAM before the Sampler's Reference Image Size setting applies.
V3.9.3 is the current source version on `main`. Download versioned source from [GitHub Releases](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases); ComfyUI Registry publication is separate. Historical Releases and tags remain available. Configured 1024×1024 Reference-only GPU/API runs succeeded, but the unchanged default template and browser save/reload remain separate acceptance gates. For a reproducible older environment, use the matching historical Release/tag and workflow; the retained V3.8 Sampler on current `main` specifically supports V3.8X2 workflows.
`pyproject.toml` declares ComfyUI `>=0.32.0`; earlier real-generation/GPU validation used ComfyUI `0.34.2`, and the scoped Fixed cache comparison above used `0.38.0`. The minimum declaration does not establish V3.9 runtime acceptance on 0.32.0. The previous Windows CPU run is `1696 passed / 1 skipped / 0 failed`; the earlier separate Linux rerun report records `1534 passed / 3 skipped / 0 failed` and does not include the new cache or external INT tests. Neither replaces browser save/reload or GPU acceptance for the unchanged official template.
## V3.9 Reference Images
### Reference Images node: choose the image for each chunk
![H3 Continuum Reference Images V3.9 node with fixed image slots and a per-chunk assignment table](images/v39/v39-reference-images-node.png)
*Example UI, not the downloadable workflow's default assignments.* Image 1–9 are fixed slots: Image 1 is always `@R1`, even when other slots are empty. With `Reference Use = Per chunk`, tick the image/chunk intersections you want; each row's `All` / `Off` buttons select or clear that image across chunks. In this screenshot, Image 1 is assigned to Chunks 1 and 2, while other images have different assignments. The official V3.9 workflow starts with its image loaders bypassed and every assignment unchecked; choose real image files, enable their loaders, then tick the intended chunks.
### Sampler V3.9: connect the helper once
![H3 Continuum Sampler V3.9 with one Reference Images Optional input](images/v39/v39-sampler-node.png)
Connect the helper's green `reference_images` output to the Sampler's single `Reference Images (Optional)` input. The V3.8X2 workflow uses its own V3.8 Sampler and older Reference sockets; opening it does not convert those connections to V3.9. This screenshot shows a **Manual 480×640 example**, not the official V3.9 workflow's saved **First Image + Draft 0.30 MP** setting. `Reference Image Size = Match Output` controls Reference conditioning size after image loading; it does not pre-resize source files or set output resolution.
V3.9 now connects all nine fixed Reference Images through one **H3 Continuum Reference Images V3.9** node. Its in-node table selects All chunks or specific chunks. The V3.9 Sampler has one typed Reference Images connection and keeps Sequence Prompt, First/Last, Review and the rest of its generation controls. Use `@R1` for Image 1; the backend converts active tags to the group's actual `<Picture N>`. Both modes use the same group-routing engine.
Prompt parsing never blocks generation. If an explicit Timeline cannot be parsed, the original text falls back to Fixed with diagnostic `H3C-P100`. An uncovered Timeline chunk reuses the previous prompt and reports `H3C-P101`. Select the image for each chunk in the Reference Images node; `@R1`–`@R9` in Sequence Prompt refer to those fixed image numbers.
For the supplied `2 × 10 s` workflow, put each time header on its own line and describe only that chunk's action. For example, after enabling the matching image loaders and checking their chunk assignments:
```text
[0-10s]
@R1 walks forward through a quiet hallway. The camera follows smoothly.
[10-20s]
@R2 enters from the right. The camera turns to follow @R2.
```
Use the actual subjects and actions from your images; this is a syntax example, not a saved prompt in the downloadable workflow. Do not place prompt text on the same line as `[0-10s]`: that is not a Timeline header. A Reference can still influence a chunk without an explicit `@R` mention when its checkbox is enabled.
The [official V3.9 workflow](../examples/workflows/MiniMax_H3_Continuum_V39.json) is the latest user-supplied JSON, copied without editing; its [single-workflow ZIP](../examples/workflows/MiniMax_H3_Continuum_V39.zip) is also available. It preserves the full V38X2 model/LoRA, Decode Cache, Finalize and Save chain. Six image loaders are prewired (initially bypassed); to use Image 7–9, add `H3 Continuum Load Image` nodes and connect them to the matching helper slots. The Sequence Prompt input remains connected but is intentionally blank: write your own prompt before Queue. The saved graph uses `Timeline`, two 10-second chunks, `Full Run` (displayed as Generate Full Video), First Image + Draft 0.30 MP size selection, and a separate Project ID. The earlier [Reference Inline migration fixture](../examples/workflows/MiniMax_H3_Continuum_V39_Reference_Inline.json) remains available unchanged. See the [step-by-step Reference Images guide](V39_REFERENCE_INLINE.md). Replace local media and model filenames as needed before Queue; enabling an image requires choosing a real file. Seven to nine active references can be particularly memory-intensive: start with fewer, smaller source images. A full ComfyUI backend restart and browser hard-refresh are required before the new node definitions appear. Browser save/reload and GPU acceptance of this unchanged default graph are still pending; a GitHub Release does not imply these gates or a Registry package have been completed.
**Compatibility:** Existing V3.8X2 workflows still use the V3.8 Sampler and old Reference Images 4–9 helper. V3.9 needs its new helper and connection. Open the new V3.9 workflow instead of only changing the Sampler type in an old graph: older V3.9 prototype workflows with direct 1–3 connections and Sampler-side routing widgets need manual rewiring, and their saved widget positions may differ. Accepted V3.8 Runs/Takes are not automatically migrated. Keep a copy of the previous workflow and use a new Run Name for V3.9.
Earlier browser checks covered the previous V3.9 controls. The new helper has GPU/API execution evidence; its browser save/reload and the unchanged default template's full acceptance are still open. CPU/JS and manifest results are separate; no V3.9 Release/Registry publication is implied.
**Version boundary:** V3.8X2 package 3.8.3 and V3.9.0 are separate workflows. This update does not create a V3.8X2 or V3.9 GitHub Release, tag, maintenance branch, or Registry version; it does not remove any older Release. The existing [`v3.8.0` tag](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/tree/v3.8.0) remains the reproducible V3.8.0 baseline.
## Loader persistence repair on main (2026-09-21, after 3.8.3)
The current paired V3.8X2 workflows now use **Core Load Audio** and **Core Load Video -> H3 Continuum Video Adapter**. The adapter has Force Rate (default 24) and IMAGE/AUDIO outputs, but **no Enable switch or file selector**. Core owns video selection/upload/save; the established duration-preserving frame selection is reused and audio is passed unchanged. To disable Video Guide, bypass the adapter or the whole video-input group; do not bypass only Load Video while leaving its required-input adapter active.
**H3 Continuum Load Image remains**, including Enable Image. Its mode observer now delegates to inherited Core setters (Issue #23), and no longer replaces the widget collection while drawing or saving. Enable is appended after Core's widgets; Core handles the filename and native Bypass persistence. The older Audio/Video node IDs remain loadable as deprecated compatibility nodes, not as the recommended new-workflow loaders. The V3.8X2 baseline registered ten node IDs including these two legacy nodes; V3.9 adds the V3.9 Sampler and Reference Images helper as separate IDs, for twelve total. V3.8 Sampler inputs, generation parameters, Decode Cache, Finalize and existing latent/storage contracts are unchanged.
After updating, restart ComfyUI and reload/refresh the frontend. A workflow that already omitted its filename or saved an incorrect mode cannot recover those missing values: reselect the intended file and Bypass state, then save once. Existing user workflows and saved Takes are not automatically rewritten. Changing upstream node types can legitimately create a new Run Storage revision; preserve the original workflow when resuming accepted runs.
Validation details and limits: [Loader repair record](LOADER_PERSISTENCE_REPAIR.md). CPU/store-model regression tests are not a substitute for the pending Windows Core 0.36.0 / frontend 1.53.6 browser tab-switch acceptance. No Release/tag or Registry publication is implied by this main-branch repair. Older loader screenshots below describe the retained legacy nodes.
## Timeline Video — Experimental on main
The existing video IMAGE input is now displayed as **`Timeline Video Frames`** and has one appended optional control, **`Video Reference Mode`**. Existing workflows and API calls that do not contain the new setting keep the old behavior through the default **`Repeat Reference`** mode.
- **`Repeat Reference`** keeps the legacy Video Guide contract: each physical generation group receives the same bounded prefix from the source. It does not loop or stretch a short source.
- **`Follow Timeline`** uses the 24 fps source interval corresponding to each physical output group. Continuation-prefix frames are not consumed twice. If the source ends, the remaining groups continue without video reference; a final usable partial interval may receive only H3's short frame-grid padding.
This setting is independent of `Prompt Format = Timeline`. It does not add a new socket or loader, does not use the source video's audio automatically, and does not provide frame-perfect motion copying. The complete decoded IMAGE batch is still held in system memory, so this feature is not a long-input RAM optimization.
Use the existing connection:
```text
Core Load Video -> H3 Continuum Video Adapter -> Timeline Video Frames
```
The two comparison workflows are intentionally separate from the unchanged official V3.8X2 workflows:
- [Follow Timeline JSON](../examples/workflows/MiniMax_H3_Continuum_V38X2_Timeline_Experimental_Follow.json) | [ZIP](../examples/workflows/MiniMax_H3_Continuum_V38X2_Timeline_Experimental_Follow.zip)
- [Repeat Reference JSON](../examples/workflows/MiniMax_H3_Continuum_V38X2_Timeline_Experimental_Repeat.json) | [ZIP](../examples/workflows/MiniMax_H3_Continuum_V38X2_Timeline_Experimental_Repeat.zip)
Local acceptance passed with full CPU `1424 passed / 1 skipped / 0 failed`, browser save/reload, and Follow/Repeat `2 x 5 s` GPU runs at 704x416, 24 fps, 240 frames, and 10-second 32 kHz stereo audio. Follow changed to the later source interval for Chunk 2 while Repeat retained the shared prefix; both completed without OOM, NaN, allocation failure, or crash. Peak observed VRAM was about 15.6–15.7 GiB on the tested RTX 5060 Ti 16 GB system, so 16 GB headroom was small. This is a functional mode-separation result, not a universal speed or subjective-quality guarantee. See [Timeline Video Experimental](TIMELINE_VIDEO_EXPERIMENTAL.md).
> **V3.8X2** is the product and workflow label for package `3.8.3`. It keeps the V3.8 production sampler and adds optional Reference Images 4–9 plus the built-in Decode Cache Helper. The older V3.8.0 package remains available from tag [`v3.8.0`](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/tree/v3.8.0).
<img width="1536" height="1024" alt="exec-9cfa73b4-c1d3-4416-8950-fa661b946638_2" src="https://github.com/user-attachments/assets/8df15764-6db6-4d4c-9eb5-0d2c7b0668f0" />
## Official V3.8X2 workflows
- **Standard name:** [JSON](../examples/workflows/MiniMax_H3_Continuum_V38X2.json) | [ZIP](../examples/workflows/MiniMax_H3_Continuum_V38X2.zip)
- **Helper-explicit name:** [JSON](../examples/workflows/MiniMax_H3_Continuum_V38X2+Decode_Cache_Helper.json) | [ZIP](../examples/workflows/MiniMax_H3_Continuum_V38X2+Decode_Cache_Helper.zip)
The two V3.8X2 JSON names contain the **same official graph**; they are distribution aliases, not different configurations. Both include the built-in Decode Cache Helper with its saved `Auto / RAM 256MB / Disk 8GB / reset_token 0` settings and the nine-reference helper connection. The standard V3.8X2 ZIP now also contains the separate V3.9 JSON, while the helper-explicit ZIP contains only its correspondingly named V3.8X2 JSON. Older V3.8X workflows remain available for existing projects.
## Nine Reference Images: memory warning
Reference Images 1–3 connect directly to the Sampler. Optional images 4–9 connect through `H3 Continuum Reference Images` and then to `Reference Images (Optional)`. Using all nine is a high-resource configuration, not the normal minimum requirement.
The acceptance run completed on an **RTX 5060 Ti 16 GB / 64 GB system** only after the nine source references were kept near **0.30 MP each**. With `512 × 608` output, `3 × 5 seconds`, Spectrum Off, LoRA Off, Sage Attention, and Balanced 22-frame continuity, Sampling used about **15.0–15.5 GiB of 16,311 MiB VRAM**. There was little 16 GB headroom.
When using many references, especially all nine, resize copies of the source images to about 0.30 MP before loading or connecting them. The Sampler still performs its normal conditioning resize, but smaller source files reduce input and preprocessing pressure. Larger references, larger output sizes, extra wrappers, or concurrent GPU work may require a higher-spec GPU and more system RAM. Nine references are optional and are not guaranteed to improve quality over three.
## Built-in Decode Cache Helper
### Important: this is not an unconditional new-generation speedup
**Manual cache clear:** normally no action is needed. Press the Helper's "キャッシュをクリア" (Clear cache) button to discard that Helper's cache on the next Queue and run native Decode again. It does not delete immediately or queue automatically. The internal `reset_token` remains serialized for workflow/API compatibility and increments once per click (wrapping to 0 at the INT limit). Creating, loading, or cloning a node does not change it. Without JavaScript, or in API workflows, change `reset_token` through the traditional numeric input.
**The Helper saves time only when unchanged latents are decoded again. It does not accelerate Sampling.** A first MISS runs native Decode and stores the result; hashing and storage can make that run slower. Generating different latents with a new random seed each time normally produces MISS, so this feature should not be expected to speed up that usage.
Concrete reuse scenarios:
- **Continue by adding chunks:** when existing chunk latents are retained and decoded again for output, cached existing chunks can HIT; newly generated chunks initially MISS.
- **Fix / partial regeneration:** unchanged chunk latents can HIT when submitted again; regenerated chunks whose latent contents changed MISS.
- **Complete / Resume output:** decoding retained latents again without resampling can benefit. Merely playing an existing video file is not a cache use case.
- **Retry from Decode onward, or change Finalize/output settings:** reuse applies when the same latents and decode-related metadata reach the Helper and Decode runs again.
- **Regenerate with a fixed seed:** reuse requires identical resulting latent contents. The same seed alone does not guarantee HIT.
The action names themselves do not enable reuse. **Each Video/Audio entry must match in latent contents, shape, dtype, decode-related metadata, and VAE identity, and its cache entry must still exist in that Helper.** If ComfyUI reuses an existing output without executing Decode, there is no additional Decode saving attributable to this Helper. Connect the Helper and select `Auto` or `RAM`. In its report, `hit` means reuse, `miss` means native Decode, and `off` / `bypass` mean no cache reuse.
Cache reuse is process-local and does not survive Python restarts. Eviction and VAE identity changes can also cause MISS. Changing mode, RAM/Disk budgets, or `reset_token` clears this implementation's Helper cache. Keeping `reset_token = 0` does not reset it on every Queue.
V3.8X2 ships `H3DecodeCacheHelper` inside this Continuum package. It remains a separate public node under `MiniMax H3/Continuum/Helpers`; no second custom-node addon is needed. The original nine IDs remain registered for compatibility; the loader migration adds H3ContinuumVideoAdapter. Sampler, Finalize, Assembly Plan, Core Decode, and saved Sampler widget/socket contracts are unchanged.
Connect Sampler `video_latents` and `audio_latents` plus their native Video/Audio VAE inputs to the Helper, then connect Helper `images` and `audio` to Finalize. Keep the Sampler `assembly_plan` directly connected to Finalize. The [official V3.8X2 workflow](../examples/workflows/MiniMax_H3_Continuum_V38X2.json) preserves its saved graph and settings. It contains no Core VAE Decode nodes, so returning to Core direct Decode requires adding those two Core nodes and reconnecting Finalize, or loading a prior Core-direct V3.8X workflow. Do not simply remove the Helper and expect automatic rewiring.
`Auto` caches full physical Video decodes on private process-local disk and small Audio decodes in bounded RAM. `Off` delegates native Decode and clears the Helper cache. Cache failure falls back to native Decode; native Decode, OOM and queue interruption are not retried. Files are not reused across Python restarts. Keep only one provider of the `H3DecodeCacheHelper` node ID: the previous standalone addon must be disabled or removed during the migration, after a verified built-in installation.
At 736×416 (0.306MP), 3×5 seconds, Euler/simple 6 steps, Turbo FL2V v1.2, a warm backend and three Video HITs, one measured configuration produced a **mechanistic estimate of about 47 seconds / 24.5% less Full Run Total**. One direct A/B pair observed 50.887 seconds / 26.46%, and two Complete/Resume re-decode pairs averaged 46.309 seconds / 79.826% less Total. These are narrow observations, not repeated universal averages or an unconditional 20%+ speed promise. The first MISS can be slower, new Sampling is not accelerated, and a changed latent or VAE identity produces a MISS.
**Historical V3.8X workflow: [JSON](../examples/workflows/MiniMax_H3_Continuum_V38x.json) | [ZIP](../examples/workflows/MiniMax_H3_Continuum_V38x.zip)** — [Published releases](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases)
✅ V3.8X on `main` includes the September 8, 2026 Review hotfix: Review Each Chunk continuation, stale `Regenerate From` state, completed-sequence extension, and Render History queue handling are repaired. Update with `git pull --ff-only origin main` (or ComfyUI Manager **Update**), restart ComfyUI, and hard-refresh the browser if the old UI remains. Existing saved Takes are preserved. Issue #13 remains a separate open long-continuation quality issue and is not part of this hotfix.
H3 Continuum is a Production Sampler for generating, reviewing, partially regenerating, and resuming long-form MiniMax H3 video without restarting the entire work. V3.8 has two product layers: **Main / Production** and **Advanced**.
## Install or update first
For a new Git installation:
```bash
cd ComfyUI/custom_nodes
git clone https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum.git
```
For an existing Git checkout:
```bash
cd ComfyUI/custom_nodes/ComfyUI-H3-Continuum
git pull --ff-only origin main
```
Restart ComfyUI after cloning or pulling. If ComfyUI Manager installed the node, use its **Update** action instead of mixing Manager updates with a second Git checkout.
To install the historical V3.8.0 package into a separate checkout instead:
```bash
cd ComfyUI/custom_nodes
git clone --branch v3.8.0 --single-branch https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum.git ComfyUI-H3-Continuum-v3.8.0
```
Use only one Continuum checkout at a time. ComfyUI Manager **Update** tracks `main`; it does not select historical tags.
Download the current V3.8X2 workflow: [JSON](../examples/workflows/MiniMax_H3_Continuum_V38X2.json) or [combined V3.8X2/V3.9 ZIP](../examples/workflows/MiniMax_H3_Continuum_V38X2.zip). The V3.8X2 JSON is one Spectrum-capable graph, also usable with [LightX2V Turbo](https://github.com/ModelTC/Minimax-H3-Turbo); both Spectrum and all Turbo LoRA entries are saved disabled. Install its external Spectrum, rgthree, and KJNodes nodes before opening it. ComfyUI-Easy-Use is not required. See [Spectrum and Turbo setup](#turbo-lora-and-spectrum-in-supplied-workflows) below.
## Prompt and skill downloads
- [V3.9 Reference Images prompt skill](../write-h3-v39-reference-prompts/SKILL.md) | [ZIP](../write-h3-v39-reference-prompts.zip) — for fixed `@R1`–`@R9` image slots and per-chunk assignments. Use it with the V3.9 Reference Images node; it is separate from the general Continuum prompt skill below.
- [LLM system prompt ZIP](../H3-Continuum-LLM-System-Prompt-v1.zip) — system instructions and reference material for Continuum prompt authoring.
- [Continuum prompt skill ZIP](../H3-Continuum-Skill-v1.zip) — general chunk-aware prompt authoring for Codex and other compatible agents.
- [Continuum Dance Director skill ZIP](../minimax-h3-continuum-dance-director.zip) — long-form choreography prompts for one primary dancer, with body and camera continuity across chunks.
These are optional prompt-authoring resources, not ComfyUI custom nodes. Extract each ZIP and follow its included instructions. They do not change the sampler or guarantee generated motion or image quality.
## You can ask an AI to read this manual
This README is intentionally detailed. You do not have to memorize it. Give its GitHub URL or this file to a web-enabled AI, describe your target video, GPU/VRAM, available First Image/audio, and whether you want to review every chunk. Ask it to use the **exact V3.8 labels** below.
Example request:
```text
Read the current ComfyUI-H3-Continuum V3.8 README. I have a 16 GB GPU and want
to create a 30-second I2VA video, reviewing every chunk. Tell me exactly which
Continuum inputs and controls to use, what must stay fixed between Queue runs,
and what output length I should expect after each action. Do not use older V3.7 UI.
```
AI products, plans, and repository-reading behavior change frequently. If a URL is not read correctly, paste or upload this README instead. Do not send private workflows, local paths, tokens, or credentials to a public AI service. The README and the UI remain the source of truth; an AI summary can still be wrong.
## V3.8 supported surface
The V3.8X2 baseline registers ten node IDs, including two deprecated loader compatibility IDs; V3.9 adds a separate Sampler and Reference Images helper. Current official workflows use Core Load Audio and Core Load Video plus the Video Adapter.
- **H3 Continuum Sampler V3.8** — the Main sampler
- **H3 Continuum Finalize** — decoded Video/Audio assembly with optional seam handling
- **H3 Continuum Load Image**
- **H3 Continuum Load Audio** (legacy compatibility; new workflows use Core Load Audio)
- **H3 Continuum Load Video** (legacy compatibility; new workflows use Core Load Video + Video Adapter)
- **H3 Continuum Video Adapter** — Core VIDEO to H3 IMAGE/AUDIO, default 24fps, no Enable
- **H3 Continuum Second Pass** — the Advanced bridge for external latent processing or upscaling
- **H3 Continuum Reference Audios** — an ordered Reference Audio 1/2/3 bundle helper
- **H3 Continuum Reference Images** — optional additional Reference Images 4–9
- **Decode Cache Helper** — the existing optional decoded-result cache
The current frontend uses `Show Advanced Settings` / `Hide Advanced Settings` to change presentation without changing generation values. Saved drafts that still contain the former frontend-only `H3 Continuum View` property are migrated to the matching collapsed/expanded view and the legacy property is removed. If the frontend extension is unavailable, the complete Python-defined interface remains visible and executable.
The official V3.8X2 workflow uses the Helper path:
```text
H3 Continuum Sampler V3.8
-> Decode Cache Helper
-> H3 Continuum Finalize
-> Create Video
-> Save Video
```
The Helper delegates cache MISS entries to native Core Video/Audio Decode. A manually wired direct Core Decode path remains supported, but it has no decoded-result cache. Save and upscaling remain external. Hi-Res Fix is not part of the V3.8X2 standard workflow; use Second Pass as the bridge when an external latent processor or upscaler is needed. External processors may change only Video LATENT spatial geometry while preserving physical groups, B/C/T, finite values, first-pass Audio, and the original Assembly Plan. Finalize accepts public `IMAGE`/`AUDIO` outputs and does not require a specific decoder class. SageAttention, Sol-Attn, and Spectrum remain external MODEL wrappers. See the [V3.8 Open Integration Contract](V38_OPEN_INTEGRATION_CONTRACT.md).
## Complete V3.8 UI reference
This section covers every user-facing control and socket on the seven public V3.8 nodes. Names such as `Power Lora Loader (rgthree)`, Spectrum, SageAttention, Core VAE Decode, and Core Save Video belong to ComfyUI or another extension. They may be used in a supplied workflow, but they are not Continuum controls.
### H3 Continuum Sampler V3.8: required graph inputs
| Input | What it does | Normal use |
|---|---|---|
| `model` | MiniMax H3 diffusion model, including any upstream MODEL wrappers or selected LoRA | Required |
| `clip` | MiniMax H3 text encoder | Required |
| `video_vae` | Encodes connected image conditioning; Continuum does not use it for final decoding | Required by the node; unused by pure T2VA |
| `sampler` | ComfyUI sampling algorithm | Required |
| `sigmas` | Noise schedule and effective step count | Required |
| `Sequence Prompt` | Complete text, list, timeline, or JSON sequence prompt | Required |
The Sampler returns six outputs: a list of `video_latents`, a list of `audio_latents`, one `assembly_plan`, a text `status`, the selected `driving_audio` when used, and a `refine_context` for the Advanced Second Pass. Raw latent lists normally go through Core Video/Audio Decode before Finalize.
### H3 Continuum Sampler V3.8: optional media inputs
| Input | Role | Important distinction |
|---|---|---|
| `first_frame` | First Image for I2VA or FL2VA | Can also supply the output aspect ratio when `Size Source = First Image` |
| `last_frame` | Optional final-image constraint for FL2VA | Can trigger a Long Terminal Merge; the final pair may become one atomic review unit |
| `reference_image_1`–`reference_image_3` | Ordered appearance, identity, subject, or scene references | They never become the implicit size source |
| `Reference Images (Optional)` | Additional images 4–9 bundled by `H3 Continuum Reference Images` | Empty inputs are skipped; encoding remains in the Sampler |
| `Video Guide Frames` | A video loader's IMAGE frame batch, applied as a persistent guide | It does not carry the source video's audio; frames are interpreted at 24 fps |
| `Driving Audio` + `Driving Audio VAE` | Original audio timeline used as native guide conditioning | The selected source audio becomes final audio; generated audio and Audio Seam are bypassed |
| `Reference Audio (Optional)` + `Reference Audio VAE (Optional)` | Legacy single conditioning-only audio reference | Generated audio remains final audio |
| `Audio References (Optional)` | Ordered bundle from `H3 Continuum Reference Audios` | Do not connect this together with the legacy single Reference Audio path |
| `Still Image Guide (Optional)` | Compatibility socket inherited from the V3.7 guide contract | Experimental; not part of the V3.8X2 official workflow |
#### First Image, Last Image, and Reference Images
![First Image enabled and Last Image bypassed](images/v38-manual/first-last-image-bypass.png)
- **First Image** establishes the opening image and is the normal visual starting point for I2VA/FL2VA. It is also the only media input that can drive `Size Source = First Image`.
- **Last Image** constrains the sequence ending. Leave it `OFF` for T2VA and ordinary I2VA. With a connected Last Image, extending or regenerating the sequence may rebuild the terminal pair.
- **Reference Images 1–9** guide identity or appearance throughout generation. Inputs 1–3 remain on the Sampler; optional inputs 4–9 are on `H3 Continuum Reference Images`, connected to the Sampler's `Reference Images (Optional)` socket. The helper only bundles images: resizing, hashing, and VAE encoding remain in the existing Sampler path. Empty slots are skipped; Picture numbers follow active slot order after any First/Last Image. Existing three-loader templates still work unchanged.
Saved workflows with the earlier direct image 4/5 inputs migrate those links into a bundle after graph loading. Migration preserves the original source links and only removes a legacy input after its new link is verified. If a source is unavailable or its destination is occupied, the legacy link is retained with a warning. Old API prompts remain accepted; supplying both legacy and bundle images for the same slot is an explicit input conflict.
![Three optional Reference Image loaders](images/v38-manual/reference-images-three-bypass.png)
Every `H3 Continuum Load Image` has `Enable Image`. `ON` loads through ComfyUI Core. `OFF` uses native node bypass and makes that optional path behave as unconnected without deleting the node or cable. First, Last, and up to nine Reference Image inputs can therefore be connected without forcing every input to be active. An empty image bundle returns no reference and does not change the zero-to-three-image sampling/reuse contract.
| Public loader | Controls | Outputs |
|---|---|---|
| `H3 Continuum Load Image` | `Enable Image`, `image` file | `IMAGE`, `MASK` |
| `H3 Continuum Load Audio` | `Enable Audio`, `audio` file | `AUDIO` |
| `H3 Continuum Load Video` | `Enable Video`, `Video` file, `Force Rate` | `images`, `audio` |
#### Video, audio, and their bypass switches
![Continuum Video and Audio loaders with Enable controls](images/v38-manual/input-video-audio-bypass.png)
- `H3 Continuum Load Video`: `Enable Video`, `Video`, and `Force Rate`. `Force Rate = 0` uses source FPS; a positive value drops/duplicates frames to that rate while preserving nominal duration and audio. Connect `images` to `Video Guide Frames` when visual video guidance is wanted. Connect `audio` to `Driving Audio` only when the source audio should guide and become the final output.
- `H3 Continuum Load Audio`: `Enable Audio` and the audio file. Use it for Driving Audio or Reference Audio according to the socket you connect.
- The node titled `Audio Switch` in the example image is ComfyUI Core's `If/Else Switch` (`ComfySwitchNode`); `Fast Groups Bypasser (rgthree)` belongs to the rgthree extension. Neither is a Continuum node or a requirement of Continuum. The Continuum loaders already provide their own native ON/OFF controls.
`Audio Continuity` is different from both audio inputs: it passes the previously generated audio context into the next generated chunk. It does not import a source track. Driving Audio replaces final generated audio; Reference Audio only conditions generation.
### Main Sampler controls
| Displayed control | Meaning | Recommended starting point |
|---|---|---|
| `Prompt Format` | `Auto`, `Fixed`, `List`, or `Timeline` interpretation of `Sequence Prompt` | `Auto` unless you need to force one parser |
| `Continuity` | Prior Video context retained at each chunk boundary | `Balanced — 22 frames` for Standard audiovisual continuation |
| `Base Seed` | Base for deterministic per-chunk seed derivation | Keep the value unchanged when continuing, comparing, or selecting Takes |
| `Control After Generate` | What ComfyUI does to `Base Seed` after a Queue | `fixed` for `Review Each Chunk`, resume, and controlled comparisons |
| `Audio Continuity` | Carries generated-audio context across boundaries | `true` for normal audiovisual generation |
| `Chunks` | Planned total chunk count, from 1 to 16 | Enter the final total, not the number to add next |
| `Seconds per Chunk` | Duration shared by every chunk | 5–15 seconds is the validated range; see duration notes below |
| `Total Length` | Read-only `Chunks × Seconds per Chunk` | Check only; it never edits either input |
| `Size Source` | `First Image` or `Manual` output geometry | First Image for I2VA/FL2VA; Manual for T2VA or exact dimensions |
| `Resolution` | First-Image sizing preset | `Draft — 0.30 MP` for tests, `Balanced — 0.60 MP` for more detail, or `Native 768` for the H3-native short edge |
| `Custom MP` | Custom First-Image pixel budget | Visible only when `Resolution = Custom` |
| `Width` / `Height` | Exact output canvas in 32-pixel steps | Visible/editable only in Manual mode; 32–16384 is accepted |
| `Run` | `Generate Full Video` or `Review Each Chunk` | Use Review when you want a human decision after each physical group |
| `Progress` | Saved raw chunks, safe resume, and Take history | Review turns it On automatically; enable it for resume/regeneration |
| `Ready to Queue` | Plain-language summary of the next Queue | Read it before pressing ComfyUI's top-right blue `Run` button |
| `Show Advanced Settings` / `Hide Advanced Settings` | Shows or hides technical controls | Visibility only; it does not change their values |
`Continuity = Balanced — 22 frames` with `Continuation Backend = Standard` and `Audio Continuity = true` uses the current Masked AV route. `Fast — 5 frames`, `Strong — 39 frames (Experimental)`, and `Auto — conservative` safely resolve to the older Reference Context route while generated-audio continuity is on. With Audio Continuity off, Standard uses the Video-only masked route.
#### Choose First Image sizing or Manual sizing
![First Image sizing with Draft 0.30 MP](images/v38-manual/sampler-size-first-image-draft.png)
`Size Source = First Image` preserves the connected First Image aspect ratio. `Resolution` then chooses its pixel area:
| Preset | Target |
|---|---|
| `Draft — 0.30 MP` | Fastest and lowest-memory First Image starting point |
| `Balanced — 0.60 MP` | More detail with greater memory and processing cost |
| `Native 768` | 768 px short-edge target with a 1344 px long-edge cap |
| `Custom` | `Custom MP` selected by the user; separate from the Native 768 cap |
![Size Source First Image tooltip](images/v38-manual/size-source-first-image-tooltip.png)
`Size Source = Manual` ignores the preset and uses exact Width/Height values. Both must be multiples of 32. Manual is the normal choice for T2VA and also supports a 32×32 diagnostic canvas when only generated audio is being checked.
![Manual Size Source with explicit Width and Height](images/v38-manual/size-source-manual-640.png)
If First Image mode is selected but no usable First Image reaches the Sampler, V3.8 uses the stored Manual Width/Height as a safe fallback and reports the dimensions in `status`. Select `Size Source = Manual` to inspect or edit those values; Width/Height are hidden in the First Image facade. Reference Images and Video Guide never become implicit size sources. The historical `Auto / Landscape / Portrait / Square` values are migration/API compatibility values, not current Main choices.
Size selection is built into the V3.8 Sampler. A separate megapixel, empty-latent, or image-size node is not required for the normal Continuum workflow.
#### Choose a chunk duration
- **5 seconds** is a valid H3 duration and the fastest practical choice for workflow checks, Review UI checks, and quick Takes.
- **8–10 seconds** usually reduces the number of handoffs while keeping each Queue easier to review and rerun.
- **15 seconds** is the [upstream MiniMax H3 native maximum per generation](https://github.com/MiniMax-AI/cli/blob/main/skill/h3-video/SKILL.md) and minimizes the number of boundaries, but it increases per-chunk time and memory. Longer is not automatically higher quality.
- Continuum accepts 4–30 seconds for compatibility and experimentation. Values above 15 seconds are not the same as an upstream native-duration recommendation and can be expensive at high resolution.
For an audio-only listening diagnostic, use T2VA, `Size Source = Manual`, and a very small canvas such as 32×32, then play the saved video and ignore its picture. This is useful for hearing music or boundary continuity; it is not evidence of normal-resolution video quality.
### Advanced Sampler controls
| Control | When it appears | Meaning |
|---|---|---|
| `Continuation Backend` | Advanced open | `Standard` is the V3.8 path; `Compatibility` restores the older Reference Context route for comparison |
| `Regenerate From` | Advanced open and Progress On | `Auto` resumes the longest compatible prefix; `Chunk N` reuses earlier chunks and rebuilds N through the target |
| `Variation Nonce` | Explicit `Regenerate From = Chunk N` | `0`: automatic variation selection with interrupted-run resume; `1` or higher: a fixed variation value. Base Seed and the other generation settings stay unchanged |
| `Run Name (Optional Override)` | Advanced open and Progress On | Stable name for resume and Render History; blank uses the Sampler's automatic identity |
| `Reference Image Size` | Advanced open and a Reference Image connected | `Match Output` is practical; `Max Identity` preserves more reference detail and may use more memory |
| `Video Guide Size` | Advanced open and Video Guide connected | `Efficient — 0.4 MP`, `Balanced — 0.6 MP`, or `Match Output` |
Frontend-managed IDs, selected Take IDs, one-shot review actions, diagnostics, preview, and legacy compatibility values are deliberately not editable as ordinary node widgets. The corresponding settings or action buttons below are the supported interface.
### Review-state controls
When `Run = Review Each Chunk`, `Progress` is On, and a review unit is ready, the settings view changes to this action view:
| Button | What the next top-right ComfyUI `Run` does |
|---|---|
| `Use it and continue` | Accepts the current result, reuses it, and generates one next physical group |
| `Try this chunk again` | Keeps the earlier accepted prefix and creates another Take of the current review unit |
| `Start again from Chunk 1` | Selects a fresh branch from Chunk 1 with current settings and automatic variation; press Queue afterward. Saved Takes are kept. The selection resets to Auto after Queue |
| `Use it and finish the rest` | Accepts the current result and generates every remaining group without further review pauses |
| `Back to Settings` | Shows the normal settings; does not queue, reset, or discard the review |
| `Return to Review` | Returns from settings to the pending review; does not undo edited values |
| `Render History — N Takes` | Opens or closes stored Take browsing; N is Take count, not chunk count |
| `Previous Take` / `Next Take` | Changes the selected stored Take only |
| `Use This Take` | Makes the selected Take canonical on the next Queue; no new sampling is required |
| `Continue From Here` | Branches from the selected Take and generates the next review unit |
The action button selects intent and gains a `✓`; it does not start generation. Press ComfyUI's top-right blue `Run` button afterward. Full step-by-step examples begin in [Review, continue, and revise a chunked video](#review-continue-and-revise-a-chunked-video).
### H3 Continuum Reference Audios
`H3 Continuum Reference Audios` bundles up to three standalone audio references behind one `Audio References (Optional)` Sampler socket. Connect the references without gaps and use `<Audio 1>`, `<Audio 2>`, and `<Audio 3>` in the same order in the prompt. One shared Core Audio VAE encodes each reference independently. This path never replaces generated final audio and does not change the separate Driving Audio contract. Existing workflows may keep using the legacy single `Reference Audio (Optional)` socket; do not connect the legacy and bundled paths together.
Its complete interface is `Reference Audio 1 (Optional)`, `Reference Audio 2 (Optional)`, `Reference Audio 3 (Optional)`, one shared `Reference Audio VAE`, and the `audio_references` output. Leave no gap in the connected order: do not connect Audio 1 and Audio 3 while leaving Audio 2 empty.
The small arrow/triangle seen beside `audio_references` is ComfyUI's connection direction/cable rendering, not an additional option. Place this helper to the left of the Sampler or use a reroute node if the cable crosses the node body.
### H3 Continuum Finalize
Finalize is normally left wired in the workflow rather than operated repeatedly. It assembles the decoded physical groups into the exact requested timeline. Required inputs are decoded `images`, decoded `audio`, and the Sampler's `assembly_plan`; optional `driving_audio` preserves the selected source track.
| Control | Meaning | Normal choice |
|---|---|---|
| `Audio Seam` | `Auto` corrects decoded audio boundaries only; `Off` leaves them unchanged | `Auto` unless diagnosing audio |
| `Video Seam` | `Auto`, experimental `Auto 2`, `Analyze Only`, or `Off` | Keep the supplied workflow value; use Analyze Only when measuring without altering frames |
| `Buffer Backend` | `Auto`, `RAM`, or `Disk-backed` storage for the assembled IMAGE buffer | `Auto` |
`Exact Total Duration` is managed as On and `Report Detail` is controlled from ComfyUI Settings. Driving Audio bypasses generated audio and Audio Seam, but Video Seam can still apply. Finalize outputs final `images`, final `audio`, and a text `report` for the downstream Save Video node.
### H3 Continuum Second Pass
Second Pass is the Advanced bridge after the complete first pass. It takes `model`, `clip`, `sampler`, low-denoise `sigmas`, externally processed `video_latents`, the matching first-pass `audio_latents`, the original `assembly_plan`, and a separate `refine_seed`. Optional `refine_context` restores the physical group's First/Last/Reference conditioning; optional `video_vae` is used only when that context requires image re-encoding.
Its outputs are `refined_video_latents`, bit-exact first-pass `audio_latents`, an `updated_assembly_plan`, and `status`. Complete Review first: Second Pass is not supported on a partial Review sequence. It is not an automatic Hi-Res button and does not include an upscaler.
### ComfyUI Settings added by Continuum
| Setting | Default | Effect |
|---|---:|---|
| `H3 Continuum: Sampling Preview` | On | Shows live sampling previews; Off only reduces preview overhead |
| `H3 Continuum: Developer Diagnostics` | Off | Enables developer-only logging/assertions; leave Off for normal use |
| `H3 Continuum: Detailed Report` | Off | Adds read-only detail to Sampler status and Finalize report; generated tensors are unchanged |
### Turbo LoRA and Spectrum in supplied workflows
![Example third-party LoRA selector with one LightX2V Turbo LoRA enabled](images/v38-manual/turbo-lora-choices.png)
The four MiniMax H3 Turbo files shown in the example come from the LightX2V MiniMax-H3-Turbo project: [GitHub](https://github.com/ModelTC/Minimax-H3-Turbo) and [Hugging Face files](https://huggingface.co/lightx2v/Minimax-h3-Turbo). `FL2VA` files are for T2VA/I2VA/first-last workflows; `Ref2VA` is for Reference-to-Video. `_768p` identifies the training-resolution family. Select **one LoRA matching the task**—do not enable all four together merely because they are listed.
| File shown | Intended path | Distilled target recorded by the file/repository |
|---|---|---|
| `minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors` | FL2VA / T2VA, 544p family | 8-step; the upstream table also lists 4-step inference as an option for this checkpoint |
| `minimax_h3_fl2v_turbo_8step_v1.0_768p_comfyui_bf16.safetensors` | FL2VA / T2VA, 768p family | 8-step |
| `minimax_h3_ref2v_turbo_8step_v1.0_768p_comfyui_bf16.safetensors` | Ref2VA, 768p family | 8-step |
| `minimax_h3_fl2v_turbo_4step_v1.2_768p_comfyui_bf16.safetensors` | FL2VA / T2VA, 768p family | 4-step |
The number in a Turbo filename is its distilled NFE target. Use the matching upstream recommendation unless a provided workflow explicitly documents a tested exception. In particular, using a 4-step LoRA with `Steps = 6` is a workflow-specific experiment, not the LightX2V default. `Euler`/`simple`, the actual Steps value, and the selected LoRA must be recorded together when comparing results.
`Power Lora Loader (rgthree)`, KJNodes SageAttention, and Spectrum are external components. Continuum does not install, enable, or tune them. The [V3.8X2 workflow](../examples/workflows/MiniMax_H3_Continuum_V38X2.json) is the supplied graph, distributed under two official filename aliases as matching JSON and ZIP pairs. Its saved default has Spectrum disabled and no Turbo LoRA enabled. These three custom-node packages are required to open the complete graph, including when switching it to Turbo. ComfyUI-Easy-Use is not required.
The supplied prompt, media selections, and node titles have deliberately been preserved. Choose files available on your computer, disable unused optional inputs, and enter your own prompt before generating. Models, LoRAs, images, and audio are not included. A custom title such as `Save 3x5s Video` is only a saved label: actual duration follows the Sampler's `Chunks` and `Seconds per Chunk`, not that title.
The same graph can be used as the quick Turbo path: disable Spectrum, enable exactly one task-matched LightX2V Turbo LoRA, then select the sampler and Steps intended for that LoRA or for a separately documented tested variation. The supplied graph lists the FL2VA Turbo 4-step v1.2 768p file but leaves it disabled in the Spectrum default. The previously tested `Euler` / `simple` / `Steps = 6` setup with that 4-step LoRA is a workflow-specific variation, not the LightX2V upstream default. Do not enable Spectrum and a Turbo LoRA together merely because both controls are present.
### Known long-continuation limitation: Issue #13
[Issue #13](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/issues/13) remains open. Repeated continuation can increase contrast, edge energy, or apparent sharpness on some prompts and trajectories. V3.8 Standard has **no active production mitigation** for this issue; it preserves the accepted Masked AV behavior.
An experimental prefix-amplitude renormalization produced a strong improvement in one `6 × 5 s` run, but a later frozen `4 × 8 s` A/B/C gate did not reproduce the target drift and therefore applied gains of `1.0`; Standard and the experimental arm were bit-exact. The older Compatibility route did not justify replacing Standard and raised a separate boundary-audio concern. The accurate release statement is therefore: **the issue is not proven fixed, the current Standard path is unchanged, and experimental mitigation stays Default Off**. Report the first and worst affected chunk, workflow JSON, prompt, seed, model/LoRA, Steps/SIGMAS, size, Continuity, and Audio Continuity when reproducing it.
> **Support boundary:** The original V3.8 launch surface had seven IDs. V3.8X2 registers ten IDs including the Video Adapter and two deprecated loader compatibility IDs; V3.9 adds two more, for twelve. Some retain earlier IDs, including Finalize (`H3ContinuumAssembleSeamV35`) and Second Pass (`H3ContinuumSecondPassV35`); this does not make all older workflows compatible. Use the matching historical Release/tag for unsupported node IDs. See [V3.8 Release and Migration Policy](V38_RELEASE_AND_MIGRATION.md).
The accepted 16GB GPU gates reached about `15.5-15.6 GiB` in the most complex cases. Actual use varies by GPU, driver, backend, model precision, resolution, and connected nodes; this is not a universal 16GB guarantee.
### Review, continue, and revise a chunked video
> **Live UI verification:** The repaired Review flow was rechecked in ComfyUI on 2026-09-07. The current walkthrough used Turbo4, actual `Steps = 4`, Euler/simple, `Draft — 0.30 MP`, and Spectrum Off. Chunk 1 reached the three-action review screen; `Use it and continue` then reused the accepted first chunk, generated only Chunk 2, and reached `Saved sequence is complete`. `Try this chunk again` was also live-validated after the Review repair. Timings below are measurements from that machine, not speed guarantees.
Start here when you already have a working V3.8 workflow and want to keep good chunks, retry a result, or extend a sequence. These instructions use the labels displayed on **H3 Continuum Sampler V3.8**. The **blue `Run` button at the top right of ComfyUI** starts execution. It is different from the Sampler's `Run` setting (`Review Each Chunk` or `Generate Full Video`). Continuum's review messages still say “press Queue”; on the frontend shown here, that means clicking the top-right blue `Run` button. Older frontends may label that button `Queue`.
![ComfyUI's top-right blue Run button starts the selected action](images/v38-manual/live-execute-button.png)
Choose the task you need:
- [Prepare the settings for continuation](#1-prepare-the-settings-for-continuation)
- [Understand the first, shorter output](#2-generate-and-review-the-first-chunk)
- [Keep the result and generate the next chunk](#3-keep-the-result-and-generate-the-next-chunk)
- [Try the current chunk again](#4-try-the-current-chunk-again)
- [Keep the result and finish everything remaining](#5-keep-the-result-and-finish-everything-remaining)
- [Go back to settings without losing the review](#6-go-back-to-settings-without-losing-the-review)
- [Choose an earlier Take or branch from it](#7-choose-an-earlier-take-or-branch-from-it)
- [Regenerate from a particular chunk](#8-regenerate-from-a-particular-chunk)
- [Add more chunks to a completed sequence](#9-add-more-chunks-to-a-completed-sequence)
- [Understand continuation time and check reuse](#10-understand-continuation-time-and-check-reuse)
- [Troubleshoot a missing action or unexpected result](#11-troubleshoot-a-missing-action-or-unexpected-result)
#### Before you begin: total length is not the next Queue's output length
`Chunks` is the **total number of chunks you want**, not the number to add on the next Queue. `Seconds per Chunk` is the duration of each chunk. `Total Length` is the planned final duration; it is not a progress counter.
For example, `Chunks = 6` and `Seconds per Chunk = 5` means a 30-second target. With `Run = Review Each Chunk`, the first Queue normally creates only Chunk 1: a 5-second video. Continuing generates the next review unit. During review, the preview/save output is the **current review unit** (normally 5 seconds), not an accumulating preview. After the run reaches `complete`, or after `Use it and finish the rest` generates the remaining chunks, Finalize outputs the assembled completed sequence.
The screenshots show **two chunks as a small example**. The same controls apply to three, four, five, six, or another supported total.
> **Last Image exception:** The examples below assume ordinary, separately generated chunks. When Last Image uses Long Terminal Merge, the final pair is generated and reviewed together. Continuing or retrying that pair operates on both chunks as one unit. Do not expect a separate review stop between them.
#### 1. Prepare the settings for continuation
1. Find `Base Seed` and keep its current number. You do not need to copy the number in the screenshot.
2. Directly below it, set `Control After Generate` to `fixed`.
![Base Seed and Control After Generate set to fixed](images/v38-manual/seed-fixed.png)
3. Set `Chunks` to your intended **final total**. Use `2` for the screenshot's example, or your own total such as `6`.
4. Set `Seconds per Chunk` to the duration you want. The example uses `5`.
5. Check `Total Length`: two chunks of 5 seconds should show `10 seconds`; six should show `30 seconds`.
6. Set `Run` to `Review Each Chunk`.
7. Set `Progress` to `On — Resume and Takes available`.
8. If `Regenerate From` is visible under `Show Advanced Settings`, leave it at `Auto` for normal continuation.
9. Check that the green `Ready to Queue` card describes the total and says `Review each chunk`.
![Review settings with fixed seed, two 5-second chunks, Review Each Chunk, and Progress On](images/v38-manual/live-review-settings.png)
**Keep these settings while continuing:** `Base Seed`, `Control After Generate = fixed`, `Seconds per Chunk`, and `Progress = On — Resume and Takes available`. Keep the same workflow, prompt, input media, model, sampler, step schedule, size, and continuation settings too. Leave `Run Name (Optional Override)` unchanged. Changing generation inputs can make saved chunks incompatible; keeping the seed alone does not guarantee reuse.
For ordinary continuation, do not increase `Chunks` after every Queue. If the goal is six chunks, leave it at `6` until those six are complete.
#### 2. Generate and review the first chunk
1. Click the top-right blue `Run` button once.
2. Wait for the workflow, including video saving, to finish.
3. Play the resulting video. In the 5-second example, the first output is **5 seconds**, even though `Total Length` says 10 or 30 seconds.
4. Look at the Sampler. While more chunks remain, it should show `Chunk 1 is ready for review` and the actions below.
![Chunk 1 ready for review with continue, retry, finish, and Back to Settings buttons](images/v38-manual/live-review-actions-steps4.png)
The shorter output is intentional: Continuum is waiting for your decision before spending time on later chunks. If your total is only one chunk, it is already complete; there is no next chunk within that target.
`Render History — 68 Takes` in this screenshot is a count of stored Takes from that run's history. It does **not** mean the target is 68 chunks. Your count will differ.
#### 3. Keep the result and generate the next chunk
1. Click `Use it and continue` on the review panel.
2. Confirm the button now shows `✓ Use it and continue`. The card says `Selected: Use it and continue. Press Queue to run this action.`
![Use it and continue selected; Queue is still required](images/v38-manual/live-continue-selected-steps4.png)
3. Click the top-right blue `Run` button. Clicking the review button alone does not start generation.
4. Wait for completion and play the new output. The accepted chunks are reused; the next chunk is newly generated.
5. If the target is not yet complete, review the new chunk and repeat from step 1.
**Example:** With `Chunks = 6` and three 5-second chunks already completed, this action reuses Chunks 1–3 and generates Chunk 4. The saved video becomes 20 seconds, and you review Chunk 4 next. It does not immediately generate Chunks 5–6.
With the screenshot's `Chunks = 2`, continuing after Chunk 1 completes the target and produces a 10-second video. There is no third chunk unless you increase the target as described in [section 9](#9-add-more-chunks-to-a-completed-sequence).
![The two-chunk target is complete after only Chunk 2 is generated](images/v38-manual/live-sequence-complete-steps4.png)
At completion, `Try this chunk again` means “make another Take of the last reviewed chunk,” here Chunk 2. It does not mean that Continuum discarded Chunk 1. Use `Back to Settings` to change the target or `Render History` to inspect stored Takes.
#### 4. Try the current chunk again
Use this when the **chunk currently being reviewed** is the one you want to replace.
1. Keep `Base Seed` unchanged and `Control After Generate = fixed`. Do not switch to `randomize` to get another Take.
2. Click `Try this chunk again` on the review panel.
3. Confirm that this action has the `✓`, then click the top-right blue `Run` button.
4. Play the new result. Continuum keeps the earlier accepted chunks and creates another Take of the reviewed chunk, with its own variation managed internally.
5. If you like it, use `Use it and continue`. Otherwise, repeat `Try this chunk again` and the top-right blue `Run` button, or select an older Take from `Render History`.
**Example:** While reviewing Chunk 3 of a six-chunk target, retrying reuses Chunks 1–2 and generates a new Chunk 3. The output is still 15 seconds, not 20 or 30 seconds. Chunks 4–6 have not been generated on this path yet. The older Take remains in history.
![Try this chunk again selected; the top-right Run button executes the retry](images/v38-manual/live-retry-selected.png)
To replace an earlier chunk after moving past it, use [Render History](#7-choose-an-earlier-take-or-branch-from-it) or [Regenerate From](#8-regenerate-from-a-particular-chunk), not the current-chunk retry button.
If you change the seed, prompt, references or other generation inputs after a review, the panel shows **Settings changed** and withdraws the old Continue/Retry/Finish actions. Queue the current settings for the backend to decide compatible reuse, or select **Start again from Chunk 1** and then Queue for an explicit restart. This is non-destructive: existing Takes remain in Render History. If you manually select `Regenerate From = Chunk N`, press Queue directly; Review buttons no longer discard that boundary.
#### 5. Keep the result and finish everything remaining
1. Play and approve the chunk currently being reviewed.
2. Leave `Chunks` at the final total you want.
3. Click `Use it and finish the rest`.
4. Confirm its `✓`, then click the top-right blue `Run` button.
5. Wait for all remaining chunks and the final video to finish. This Queue does not pause for a review after each remaining chunk.
These three buttons are different even though all require `Queue`:
| While reviewing Chunk 3 of a six-chunk, 5-second-per-chunk target | Reused chunks | Newly generated chunks on the next Queue | Saved duration |
|---|---|---|---|
| `Use it and continue` | 1–3 | 4 | 20 seconds |
| `Try this chunk again` | 1–2 | A new Take of 3 | 15 seconds |
| `Use it and finish the rest` | 1–3 | 4–6 | 30 seconds |
For a two-chunk target after Chunk 1, `Use it and continue` and `Use it and finish the rest` produce the same remaining amount: only Chunk 2 is left. Their difference becomes clear with a larger target.
#### 6. Go back to settings without losing the review
1. On the review panel, click `Back to Settings`.
2. The normal controls appear again. You can inspect `Chunks`, `Run`, `Progress`, and the other settings.
3. Click `Show Advanced Settings` only if you need controls such as `Regenerate From` or `Run Name (Optional Override)`.
4. To go back without changing the operation, click `Return to Review`.
![Back to Settings keeps the completed sequence and exposes Return to Review](images/v38-manual/live-settings-complete-steps4.png)
![Return to Review above Regenerate From, Run Name, and Continuation Backend](images/v38-manual/return-to-review.png)
**Neither navigation button queues work, deletes a Take, resets the run, nor cancels a selected action.** `Return to Review` also does not undo settings you edited. Before the next Queue, select the action you actually want and check its `✓`.
`Show Advanced Settings` and `Hide Advanced Settings` only show or hide controls; they do not turn those settings on or off. `Back to Settings` is not a “start over” button.
#### 7. Choose an earlier Take or branch from it
A **Take** is a stored version of a generated result. A **branch** is an alternative continuation from a chosen result; the old history is retained.
1. On the review panel, click `Render History — … Takes`.
2. Use `Previous Take` and `Next Take` to choose a stored result. Read the displayed chunk and branch information; selecting an entry by itself does not replace your current sequence.
![Render History with Take navigation and the two explicit Take actions](images/v38-manual/live-render-history-steps4.png)
3. Choose one of the following actions, then click the top-right blue `Run` button:
- `Use This Take`: make the selected stored result the active result. This action runs no new sampling; the output path can still decode and save the selected sequence.
- `Continue From Here`: keep the sequence through the selected Take and generate the next review unit on a new continuation path.
4. Review the resulting output before continuing again.
**Example:** To keep an older Chunk 2 and make a new Chunk 3, select that Chunk 2 Take and use `Continue From Here`, then the top-right blue `Run` button. Do not select Chunk 3 if Chunk 3 is the part you want replaced. Later chunks from the old path are not automatically attached to the new one.
Keep `Regenerate From = Auto` when using Take actions. Do not combine an explicit `Regenerate From` selection with `Use This Take` or `Continue From Here`. If the selected Take already completes the configured total, `Continue From Here` has no next chunk to generate.
#### 8. Regenerate from a particular chunk
Use this when you know the **first chunk that must be redone**. Unlike retrying the current review, this can rebuild a longer section of an existing sequence.
1. Click `Back to Settings` if you are on the review panel.
2. Keep `Base Seed`, the total `Chunks`, and `Seconds per Chunk` unchanged. Keep `Control After Generate = fixed` and `Progress = On — Resume and Takes available`.
3. Click `Show Advanced Settings`.
4. Set `Regenerate From` to the first chunk to redo, such as `Chunk 3`. Earlier chunks can be reused only if their inputs remain compatible.
5. Set `Variation Nonce`, which appears for explicit regeneration. Keep `Base Seed` fixed. `0` selects a new variation automatically after a completed regeneration; a compatible interrupted regeneration resumes its existing variation instead. A value of `1` or higher fixes the variation: keep it to reproduce/resume that variation, or change it to request another. Repeating the same positive value with identical inputs does not request a different result. This is separate from `Try this chunk again`, which advances the review variation automatically.
![Regenerate From exposes Variation Nonce for an explicit regeneration](images/v38-manual/live-regenerate-from-chunk2.png)
6. Set `Run` to `Generate Full Video` to rebuild everything from the selected chunk through your target in one Queue.
![Run set to Generate Full Video while Progress stays On](images/v38-manual/full-video-progress.png)
7. Click the top-right blue `Run` button, wait for completion, and check the new full-length output.
8. Before ordinary continuation or review actions, set `Regenerate From` back to `Auto` so that the next Queue is not another explicit request to redo that section.
**Example:** With a completed six-chunk sequence, `Regenerate From = Chunk 3` and `Run = Generate Full Video` reuse compatible Chunks 1–2 and regenerate Chunks 3–6. This is **four newly generated chunks**, not only Chunk 3. Later chunks depend on the changed continuation and must follow the new path.
If you want a review pause instead, choose `Run = Review Each Chunk` in step 6. The next Queue generates the first requested review unit, not the whole remaining section. After reviewing it, select `Use it and continue` to proceed one unit at a time, or `Use it and finish the rest` to complete the section. Keep earlier-chunk inputs unchanged if you want those chunks reused.
#### 9. Add more chunks to a completed sequence
**You do not need Render History just to make a completed video longer.** For an ordinary extension, go back to the normal settings, increase the target `Chunks`, and Queue again. `Chunks` is the **final total number of chunks**, not the number you want to add next.
For the simplest and safest extension, keep the existing generation conditions unchanged: the same Sampler/workflow and run identity, `Base Seed`, `Control After Generate = fixed`, `Seconds per Chunk`, prompt, model/LoRA, media, dimensions, sampling schedule, Continuity, and Audio Continuity. Keep `Progress = On — Resume and Takes available` and `Regenerate From = Auto`.
**Example: extend a completed 2 × 5-second sequence to 3 × 5 seconds**
1. Finish the original two chunks. The completed target is `Chunks = 2`, so the saved video is 10 seconds.
2. Click `Back to Settings`. You do **not** need to open `Render History`.
3. Change `Chunks` from `2` to `3`. `Total Length` changes from 10 seconds to 15 seconds.
4. Keep `Progress = On — Resume and Takes available` and `Regenerate From = Auto`. Keep the prompt and the existing generation inputs unchanged for this ordinary extension.
5. Use `Run = Review Each Chunk` if you want to review the new chunk, or `Run = Generate Full Video` if you want all currently missing chunks generated in one Queue.
6. Click the top-right blue `Run` button.
With compatible saved progress, the expected work is:
- Chunk 1 → reused; no new Sampling
- Chunk 2 → reused; no new Sampling
- Chunk 3 → newly generated
- final saved sequence → 3 chunks / 15 seconds
The same rule applies to a larger extension. For example, changing `Chunks` from `3` to `5` means the new target is five total chunks; it does **not** mean “add five.” Review mode first generates Chunk 4, then Chunk 5 on the next accepted continuation. Full-video mode generates the missing Chunks 4–5 in one Queue.
Switching `Generate Full Video` ↔ `Review Each Chunk` by itself does **not** mean restart from Chunk 1. A Chunk 1 restart happens only when you explicitly choose `Start again from Chunk 1`, explicitly set Advanced `Regenerate From = Chunk 1`, or the backend determines that the stored prefix is incompatible with the current generation contract.
`Render History` is for a different job: browsing stored Takes, selecting an older Take, or branching with `Continue From Here`. Normal append-style extension does not require History. `Previous Take` / `Next Take` only browse stored results; they do not generate a new chunk.
**Do not assume append-only reuse with Last Image.** Extending a sequence changes where its final-image constraint belongs. The former last chunk, or a terminal pair, may need regeneration. Likewise, changing the prompt or other generation-contract inputs can invalidate saved progress. `Progress = On` enables compatibility checking; it does not force incompatible chunks to be reused.
The extension path is covered by CPU compatibility tests, including two-to-three chunks and Last Image invalidation. It is not the same GPU test as the two-chunk review timing example below; arbitrary extensions and media combinations have not all been GPU-verified.
#### 10. Understand continuation time and check reuse
Continuing does not mean saving only the new segment. With compatible saved progress, the normal workflow:
1. Reuses the accepted chunks' stored Video/Audio latents without sampling them again.
2. Samples the requested new chunk or remaining section.
3. Decodes the completed sequence so far, runs `H3 Continuum Finalize`, and saves the combined video again.
The full-output work grows as the sequence gets longer. Consequently, the next Queue can take as long as the first, or longer, even though earlier chunks are **not** being sampled again. Loading, preparation, decoding, assembly, saving, and other workflow overhead all contribute to total time.
The latest live GPU check used Turbo4, Euler/simple, actual `Steps = 4`, Spectrum Off, and `Draft — 0.30 MP` with `Chunks = 2`, `Seconds per Chunk = 5`, and Review enabled:
| Queue | Sampler work | H3 Continuum Sampler node time | Saved video |
|---|---|---:|---:|
| First | Generate Chunk 1 | 103.802 s | 5 seconds |
| After `Use it and continue` | Reuse Chunk 1; generate Chunk 2 | 43.738 s | 10 seconds |
Each Queue sampled one chunk with four steps. The second Queue did not sample Chunk 1 again; it sampled only the requested next chunk, then decoded, finalized, and saved the combined 10-second sequence. The first Queue also included cold-start/model preparation effects, so the difference must not be treated as a general speed ratio. Full-sequence decode and save work still occurs on continuation.
To inspect your own run:
1. Open ComfyUI Settings and search for `H3 Continuum: Detailed Report`.
2. Enable that setting before the Queue you want to inspect.
3. After completion, inspect the Sampler's `status` output through a text-display node in your workflow.
4. In the two-chunk example's second Queue, look for `1 reused, 1 generated, 2 total`. Detailed fields should show `sampled_physical_groups=1` and, with four steps, `sampling_steps_total=4`, not `8`.
The report field names are diagnostic text, not controls you need to set. For a larger target, use the reported reused/generated counts rather than timing alone. Terminal Merge can combine two chunks into one sampled group, so its group count differs from its chunk count.
#### 11. Troubleshoot a missing action or unexpected result
| What you see | What to check or do |
|---|---|
| A 5-second video despite a longer `Total Length` | In `Review Each Chunk`, this is the expected first output. Review it, select an action, then click the top-right blue `Run` button. |
| No next-chunk action | Check whether the configured total is already complete. Otherwise, wait for Queue completion and verify `Run = Review Each Chunk` and `Progress = On — Resume and Takes available`. If you opened settings, click `Return to Review`. |
| No review panel even though an incomplete Review run finished successfully | Inspect the Sampler's `status` and any ComfyUI error. This is not a reason to assume all chunks finished or to change `Base Seed`; the frontend must receive the saved review state. |
| Clicking an action appears to do nothing | Check its `✓`, then click ComfyUI's top-right blue `Run` button. The action button only selects what Queue will do. |
| `Set Control After Generate to fixed` | Set that exact control to `fixed` and retain the original `Base Seed` used for the saved chunks. |
| Settings disappeared behind the review panel | Click `Back to Settings`. Use `Return to Review` to come back; neither navigation button resets the run. |
| The next Queue takes as long as the first | Check reused/generated counts. Full-sequence decoding and saving still run; elapsed time alone cannot tell you whether chunks were resampled. |
| Earlier chunks are generated again | Check `Progress`, saved files, run identity, `Base Seed`, and changed generation inputs. Check that `Regenerate From` is `Auto` unless you intentionally requested regeneration. |
| Retrying changed more than one chunk | A terminal pair is one review unit. Also distinguish `Try this chunk again` from `Regenerate From`, which rebuilds the chosen chunk and its following section. |
| You want to stop choosing an action after every chunk | At the review panel, select `Use it and finish the rest`, then click the top-right blue `Run` button. Keep `Chunks` at your intended final total. |
Driving Audio can be used with Review and Smart Regenerate. The Continuum Image, Audio, and Video loaders support native bypass for optional media paths.
> **Second Pass limitation:** Finish the reviewed sequence before running Second Pass. Second Pass / `refine_context` is not supported on a partial Review sequence.
+309
View File
@@ -0,0 +1,309 @@
# H3 Continuum — 日本語の詳細ガイド
[日本語README](../README_JA.md) · [英語の図解ガイド](USER_GUIDE.md) · [日本語の履歴資料](README_HISTORY_JA.md)
この資料は、正本`84899c1`(パッケージ3.9.8)の日本語READMEから、詳しい操作と画像を移したものです。
V3.8の表示名・操作説明は保持しているV3.8X2用であり、V3.9の導入・Reference説明とは区別してください。
過去の版表記、修正記録、測定条件は当時の記録として保持しています。
現在の配布・互換性・制約は[短縮版README](../README_JA.md)と[KJ対応条件](KJ_PREVIEW_OVERRIDE_V398.md)を参照してください。
- [V3.9とV3.8X2の導入](#最初にv39とv38x2)
- [Reference Images](#v39-reference-images)
- [Review・継続の手順](#10秒を5秒ずつ確認しながら作る)
- [Prompt・スキル配布](#プロンプトスキルのダウンロード)
<!-- Preserved README sections begin. -->
## 最初に:V3.9とV3.8X2
ComfyUIの **Templates → ComfyUI-H3-Continuum** には、今回選んだ[V3.9公式Workflow](../examples/workflows/MiniMax_H3_Continuum_V39.json)と、別の[V3.8X2 Workflow](../examples/workflows/MiniMax_H3_Continuum_V38X2.json)を2件表示します。[V3.8X2 Workflow ZIP](../examples/workflows/MiniMax_H3_Continuum_V38X2.zip)には、**V3.8X2とV3.9の両JSON**を同梱しました。[V3.9単独ZIP](../examples/workflows/MiniMax_H3_Continuum_V39.zip)も残します。いずれもWorkflow用ZIPで、カスタムノード本体のインストーラーではありません。V3.8 Samplerは残るので既存のV3.8X2 Workflowは使えます。ただしV3.9へReference配線や保存済みRun/Takeは自動移行しません。旧Workflowを保存しておき、新しい作業ではV3.9専用Workflowを開いてください。
選択したV3.9テンプレートは32ノードです。画像Loader 1~6は新ヘルパーへ配線済み・初期Bypassです。7~9番の入力欄は残し、必要な人が画像Loaderを追加して接続します。使用する画像を選んで必要なLoaderだけ有効にしてください。保存済みサイズ設定は**First Image+Draft — 0.30 MP**で、別のFirst Image Loaderは有効です。Queue前に自分のFirst Imageファイルを選び直してください。保存済みのWidth/Height 480×640はManual時だけ有効です。`Reference Image Size`は`Match Output`で、読み込み後のconditioningを縮小しますが元画像ファイル自体は縮小しません。チャンク別Referenceの効果を明瞭に試すなら、First ImageとLast ImageをOFFにし、Manualへ切り替えてReference Imagesだけを使うのが勧められます。First/Lastを入れるとそれ自体が動画を拘束し、Last Imageは最後の2つの論理チャンクを1つの生成groupに結合する場合があります。その場合、両チャンクの有効Reference集合は同じにしてください。Referenceを外しても、前チャンクの映像から人物が引き継がれることはあります。
ヘルパーの `All chunks` は接続画像を全チャンクに、`Per chunk` は画像×チャンクの表で指定します。このテンプレートは`Per chunk`で全チェックOFFから始まるため、Loaderを有効にした後、使用したいチャンクにもチェックしてください。空欄があっても番号はずれず、Image 1は常に`@R1`、Image 9は`@R9`です。Sequence Promptにはチャンクごとの動作を短く書き、そのチャンクで使う画像の`@R`タグを付けます。内部の`<Picture N>`番号はgroupごとに自動変換されます。使わないタグは警告しますが生成を止めません。まず短いチャンクと縮小済み画像で試してください。入力元画像はSamplerのReference Image Size設定より前にRAMを使うことがあります。
V3.9.3は`main`の現行ソースです。バージョンを固定したソースは[GitHub Releases](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/releases)から取得してください。ComfyUI Registryへの公開は別の作業です。過去のReleaseとtagも残します。設定済み1024×1024のReference-only GPU/APIテストは通っていますが、無改変の公式テンプレート初期設定とブラウザー保存・再読込は別の受入項目です。古い環境を厳密に再現するには対応する過去のRelease/tagとWorkflowを使用してください。現行mainにもV3.8 Samplerを残し、V3.8X2 Workflowを使えるようにしています。
`pyproject.toml`の宣言上はComfyUI `>=0.32.0`ですが、以前の実生成・GPU検証はComfyUI `0.34.2`、上記の限定したFixedキャッシュ比較は`0.38.0`で行いました。最低宣言版でV3.9の実機動作を確認済みという意味ではありません。以前のWindows CPU試験は`1696 passed / 1 skipped / 0 failed`、以前の別環境Linux再実行報告は`1534 passed / 3 skipped / 0 failed`で、新しいキャッシュ/外部INT試験は含みません。これらは無改変の公式テンプレートの実ブラウザー保存・再読込やGPU受入を代替しません。
## V3.9 Reference Images
### Reference Imagesノード:画像とチャンクを選ぶ
![H3 Continuum Reference Images V3.9ノードの固定スロットとチャンク別割り当て表](images/v39/v39-reference-images-node.png)
*この画像は操作例で、配布Workflowの初期割り当てではありません。* Image 1~9は固定スロットで、途中が空でもImage 1は常に`@R1`です。`Reference Use = Per chunk`にすると、使いたい画像とチャンクの交点をチェックできます。各行の`All`/`Off`は、その画像を全チャンクで選択/解除します。画像ではImage 1をChunk 1・2に割り当て、ほかの画像は別のチャンクに割り当てています。公式V3.9 Workflowは画像Loaderが初期Bypass、割り当てが全OFFです。実ファイルを選び、Loaderを有効化してから、使うチャンクにチェックしてください。
### Sampler V3.9:ヘルパーから1本で接続
![Reference Images Optional入力を1本持つH3 Continuum Sampler V3.9](images/v39/v39-sampler-node.png)
ヘルパーの緑色の`reference_images`出力を、Samplerの`Reference Images (Optional)`入力へ1本つなぎます。V3.8X2 Workflowは旧V3.8 Samplerと従来のReference接続を使用し、開くだけでV3.9配線へ変換されるわけではありません。この画像のサイズ設定は**Manual 480×640の操作例**で、公式V3.9 Workflowの保存値は**First Image+Draft 0.30 MP**です。`Reference Image Size = Match Output`は画像読込後のReference conditioningサイズを決め、元ファイルを事前縮小したり、動画の出力解像度を直接決めたりはしません。
V3.9では、固定番号の参照画像1~9を新しい **H3 Continuum Reference Images V3.9** ノードにまとめます。ノード内の表で「All chunks」またはチャンク別のチェックを選び、Samplerへは1本だけ接続します。`Sequence Prompt`で`@R1`と書けば固定の画像1を指します。有効なタグは各生成groupの実際の`<Picture N>`へ変換されます。両モードとも同じチャンク別処理を通ります。
Promptは実行を止めません。明示的なTimelineでも解析できなければ元の入力をFixedとして扱い、診断`H3C-P100`を出します。Timelineで指定のないチャンクは前のPromptを再利用し、`H3C-P101`で知らせます。どの画像を各チャンクに使うかはReference Imagesノードで選び、Prompt中の`@R1`~`@R9`は接続画像の固定番号に対応させてください。
同梱の`2 × 10秒`Workflowでは、時間見出しを必ず単独行にし、各チャンクの動作だけを短く書きます。対応する画像LoaderをONにしてチャンク欄へチェックを入れた場合の書式例です。
```text
[0-10s]
@R1 walks forward through a quiet hallway. The camera follows smoothly.
[10-20s]
@R2 enters from the right. The camera turns to follow @R2.
```
人物や動作は実際の画像に合わせて書き換えてください。これは書式例で、配布WorkflowのPromptは空欄です。`[0-10s]`と本文を同じ行には書かないでください。チェックされた画像は、Prompt中に`@R`を書かなくてもそのチャンクのconditioningに入ります。
[公式V3.9 Workflow](../examples/workflows/MiniMax_H3_Continuum_V39.json)は、今回指定された添付JSONを無改変で採用しました。[単独ZIP](../examples/workflows/MiniMax_H3_Continuum_V39.zip)も用意しています。V38X2のモデル・LoRA、Decode Cache、Finalize、Saveのフル構成を維持し、画像Loaderは1~6だけを配線済み・初期Bypassにしています。7~9番を使う場合は`H3 Continuum Load Image`を追加し、ヘルパーの同じ番号の入力へ接続してください。Sequence Promptの入力ノードと配線は残し、文章は空欄です。Queue前に自分のプロンプトを入力してください。保存設定は`Timeline`、2×10秒、`Full Run`(画面表示はGenerate Full Video)、First Image+Draft 0.30 MP、別Project IDです。従来の[V3.9 Reference Inline移行用Workflow](../examples/workflows/MiniMax_H3_Continuum_V39_Reference_Inline.json)は変更せず残します。チャンク別割り当て、First/Lastとの終端結合、保存・再読込、画像メモリの注意点は[操作ガイド](V39_REFERENCE_INLINE.md)を参照してください。Queue前にモデル・素材のファイル名を確認し、画像を有効化する場合は実ファイルを選択してください。7~9枚の参照は特に高負荷なので、最初は枚数と元画像サイズを抑えてください。新ノードを表示するにはComfyUI本体の完全再起動とブラウザーの強制再読込が必要です。無改変の公式テンプレートの実ブラウザー保存・再読込/GPU受入はまだ未実施です。GitHub Releaseの公開は、これらの受入やRegistry公開の完了を意味しません。
**互換性:** V3.8X2の既存Workflowは旧V3.8 Samplerと旧4~9ヘルパーで引き続き利用します。V3.9では新ヘルパーへの接続が必要です。旧Workflow上でSamplerの型だけ変えず、上記のV3.9専用Workflowを開いてください。以前のV3.9試作WorkflowにあったSampler直結1~3やSampler側の割り当て欄は配線と保存widgetの位置が異なります。V3.8の途中Run/Takeは自動移行しません。元Workflowを保存し、V3.9では別Run Nameを使ってください。
以前のV3.9操作系のブラウザー確認は、新しいヘルパーの保存・再読込確認とは別です。新ヘルパーは設定済みGPU/API実行の証拠がありますが、公式テンプレート初期設定の受入は別です。CPU/JS・Manifestも別に判定し、V3.9 GitHub Release/Registry公開はまだ行っていません。
**バージョン境界:** V3.8X2(package 3.8.3)とV3.9.0は別Workflowです。今回、新しいtag・GitHub Release・保守ブランチ・Registry版は作成せず、既存Releaseも削除しません。旧V3.8.0は既存の`v3.8.0` tagから取得できます。
## mainのLoader保存・復元修正(2026-09-21、3.8.3以降)
現行V3.8X2ワークフローは、音声を **Core Load Audio**、動画を **Core Load Video → H3 Continuum Video Adapter** に変更しました。Video Adapterは24fpsへの変換(Force Rate初期値24)とIMAGE/AUDIO出力だけを担当し、**Enableやファイル選択UIはありません**。動画を使わない場合はAdapter、または動画入力グループ全体をBypassしてください。Load VideoだけをBypassして必須入力のAdapterをONに残す構成は避けてください。
画像は **H3 Continuum Load ImageとEnable Imageを維持**します。Issue #23の継承されたmode setterへの委譲と、描画・保存中のwidget配列差し替え廃止を適用します。EnableはCoreの欄の後ろに配置し、画像ファイルとBypassの保存はCoreに委ねます。旧Audio/VideoノードIDは既存workflow用の非推奨互換ノードとして残します。V3.8X2の登録IDは互換用2個を含め10個で、V3.9では別IDのV3.9 SamplerとReference Imagesヘルパーを追加して計12個です。V3.8 Sampler設定・Sampling・Decode Cache・Finalize・既存Run Storage契約は変更しません。
更新後はComfyUI再起動と画面の再読み込みが必要です。既に保存JSONから消えたファイル名やOFF状態は推測で復元できないため、一度選び直して保存してください。既存ユーザーworkflowや保存済みTakeは自動変更しません。上流ノードの種類変更でRun Storageが新revisionになることはあるため、過去Runの再開には元workflowを保管してください。
検証範囲は[Loader修正記録](LOADER_PERSISTENCE_REPAIR.md)を参照してください。CPU回帰テストと実ブラウザ確認は別です。Windows Core 0.36.0 / frontend 1.53.6のタブ切替確認は未実施で、Release/tag・Registry公開も別作業です。以下の旧Loader画像は互換ノードの説明です。
## Timeline Video — main上のExperimental機能
従来の動画IMAGE入力を**`Timeline Video Frames`**と表示し、末尾に省略可能な設定**`Video Reference Mode`**を1項目追加しました。新しい設定を持たない既存Workflow/APIは、初期値の**`Repeat Reference`**で従来動作を維持します。
- **`Repeat Reference`**:従来のVideo Guide契約です。各physical生成groupで、同じ先頭側の上限付き参照を使います。短い素材を自動ループ/Stretchしません。
- **`Follow Timeline`**:各physical出力groupの時間位置に対応する24fps参照区間を使います。Continuationの過去prefixを二重消費しません。素材終了後は動画参照なしで続行し、利用可能な終端区間にはH3 frame-grid整列分だけ短いpaddingを行う場合があります。
これは`Prompt Format = Timeline`とは別の設定です。新しいsocketやLoaderは増やさず、元動画のAudioも自動使用しません。参照動画はsoft conditioningであり、motionをframe単位で完全コピーする機能ではありません。また、decode済みIMAGE batch全体をsystem RAMへ保持するため、長尺入力のRAM削減機能ではありません。
接続は既存のままです。
```text
Core Load Video → H3 Continuum Video Adapter → Timeline Video Frames
```
公式V3.8X2 Workflow本体は変更せず、比較用Experimental Workflowを別名で追加しています。
- [Follow Timeline JSON](../examples/workflows/MiniMax_H3_Continuum_V38X2_Timeline_Experimental_Follow.json)/[ZIP](../examples/workflows/MiniMax_H3_Continuum_V38X2_Timeline_Experimental_Follow.zip)
- [Repeat Reference JSON](../examples/workflows/MiniMax_H3_Continuum_V38X2_Timeline_Experimental_Repeat.json)/[ZIP](../examples/workflows/MiniMax_H3_Continuum_V38X2_Timeline_Experimental_Repeat.zip)
ローカル受入ではFull CPU `1424 passed / 1 skipped / 0 failed`、ブラウザ保存・再読込、Follow/Repeatの`2 × 5秒` GPUテストがPASSしました。両方とも704×416、24fps、240フレーム、映像10秒、32kHz stereo音声10秒で、OOM/NaN/allocation failure/crashはありません。FollowはChunk 2で後半区間へ切り替わり、Repeatは同じ先頭側参照を維持しました。RTX 5060 Ti 16 GBで観測した最大VRAMは約15.6~15.7 GiBで、16 GBの余裕は小さい条件です。これは機能・mode分離の受入結果であり、一般的な速度や主観画質の保証ではありません。詳細は[Timeline Video Experimental](TIMELINE_VIDEO_EXPERIMENTAL.md)を参照してください。
> **V3.8X2**はpackage `3.8.3`の製品名・Workflow名です。V3.8 Production Samplerを維持し、任意のReference Image 4~9と内蔵Decode Cache Helperを追加しています。旧V3.8.0は[`v3.8.0`タグ](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/tree/v3.8.0)から導入できます。
## V3.8X2 公式Workflow
- **標準名:** [JSON](../examples/workflows/MiniMax_H3_Continuum_V38X2.json)/[ZIP](../examples/workflows/MiniMax_H3_Continuum_V38X2.zip)
- **Helper明示名:** [JSON](../examples/workflows/MiniMax_H3_Continuum_V38X2+Decode_Cache_Helper.json)/[ZIP](../examples/workflows/MiniMax_H3_Continuum_V38X2+Decode_Cache_Helper.zip)
V3.8X2の2つのJSON名は**同じ公式graph**を収録した配布上の別名で、異なる設定のWorkflowではありません。どちらも内蔵Decode Cache Helperと9枚Reference Image用の接続を含み、Helperの保存値は`Auto / RAM 256MB / Disk 8GB / reset_token 0`です。標準名のV3.8X2 ZIPには別のV3.9 JSONも同梱し、Helper明示名ZIPは対応するV3.8X2 JSONだけを収録します。既存projectでは従来のV3.8X Workflowも引き続き使用できます。
## Reference Image 9枚使用時のメモリ注意
Reference Image 1~3はSamplerへ直接接続し、任意の4~9は`H3 Continuum Reference Images`を経由して`Reference Images (Optional)`へ接続します。9枚すべてを使う構成は高負荷であり、通常の最低要件ではありません。
受入テストでは、**RTX 5060 Ti 16 GB/RAM 64 GB**環境で、9枚の元画像を各**約0.30 MP**に抑えた条件で完走しました。`512 × 608`出力、`3 × 5秒`、Spectrum Off、LoRA Off、Sage Attention、Balanced 22-frame continuityで、Sampling中のVRAMは`16,311 MiB`中の約**15.0~15.5 GiB**でした。16 GB環境では余裕がほとんどありません。
多数、特に9枚を使う場合は、読み込み・接続前に元画像のコピーを約0.30 MPへ縮小してください。Sampler内の通常のconditioning用Resizeは引き続き行われますが、小さい入力ファイルにすることで読み込み・前処理時の負荷を抑えられます。高解像度参照、大きな出力、追加wrapper、同時GPU処理では、より高性能なGPUと多いシステムRAMが必要になる可能性があります。9枚は任意であり、3枚より常に高画質になる保証はありません。
## 内蔵 Decode Cache Helper
### 重要:新規生成が一律に速くなる機能ではありません
**キャッシュの手動クリア**:普段は操作不要です。Helperの「キャッシュをクリア」ボタンを押すと、次回QueueでこのHelperのキャッシュを破棄し、通常Decodeからやり直します。ボタンは即時削除や自動Queueを行いません。内部の`reset_token`は保存・API互換用に残り、押すたびに1増えます(INT上限では0へ戻ります)。作成・読込・複製だけでは変更しません。JavaScriptが使えない環境やAPI版では、従来どおり`reset_token`の値を変更してください。
**短縮するのは、変更していないlatentを再びデコードする時間です。Sampling(生成計算)自体は高速化しません。** 初回は通常Decodeとキャッシュ保存を行うためMISSとなり、保存・照合の負荷で遅くなる場合があります。毎回ランダムシードで異なるlatentを新規生成する使い方では、基本的にMISSとなり、この機能による短縮は期待できません。
具体的に再利用できる場面:
- **Continueでチャンクを追加**:既存チャンクのlatentを保持し、出力時にもう一度Decodeする場合、保存済みの既存部分だけを再利用します。追加チャンクは初回MISSです。
- **Fix/部分再生成**:変更していないチャンクのlatentがそのまま再入力される場合、その部分だけを再利用します。再生成してlatentが変わった部分はMISSです。
- **Complete/Resumeで再出力**:Samplingをやり直さず、保存済みlatentから全体を再Decodeする場合に有効です。単なる既存動画ファイルの再生ではありません。
- **Decode以降のリトライ、Finalizeや出力設定の変更**:Helperへ同じlatent・Decode関連情報が再入力され、Decodeが再実行される場合に有効です。
- **固定シードで同条件を再生成**:生成後のlatentが完全に一致した場合だけ再利用できます。同じシードだけではHITを保証しません。
これらの操作名そのものが有効条件ではありません。**各Video/Audioの入力latentの内容・shape・dtype、Decode関連メタデータ、VAE識別情報が一致し、そのHelperにキャッシュが残っていること**がHITの条件です。ComfyUIが上流/出力をそのまま再利用してDecode自体を実行しない場合、Helperによる追加のDecode短縮はありません。Helperを接続し、`Auto`または`RAM`で使ってください。レポートの`hit`は再利用、`miss`は通常Decode、`off`/`bypass`はキャッシュ再利用なしを表します。
キャッシュはPythonプロセス内の利用に限られ、再起動後は再利用しません。上限による退避・破棄やVAE識別情報の変化でもMISSになります。モード、RAM/Disk予算、`reset_token`を変更すると、この実装ではHelperキャッシュをクリアします。`reset_token = 0`を維持することは、毎回リセットする意味ではありません。
`H3DecodeCacheHelper`をContinuumパッケージ内に収録し、UI上は`MiniMax H3/Continuum/Helpers`の独立公開ノードとして残します。インストール対象はContinuumだけです。V3.8X2の公開IDは、旧互換Loader 2件を含め計10件です。V3.9は別IDのSamplerとReference Imagesヘルパーを追加して計12件です。V3.8 Sampler、Finalize、Assembly Plan、Core Decode、保存済みwidget/socket契約は変更しません。
Samplerの`video_latents`/`audio_latents`と同じVideo/Audio VAEをHelperへ接続し、Helperの`images`/`audio`をFinalizeへ接続します。Samplerの`assembly_plan`はFinalizeへ直接接続したままです。[公式V3.8X2 Workflow](../examples/workflows/MiniMax_H3_Continuum_V38X2.json)は保存済みgraphと設定を維持します。このgraphにはCore VAE Decodeノードがないため、Core直結へ戻すにはCore Video/Audio Decodeを追加してFinalizeへ再配線するか、旧V3.8XのCore直結graphを開いてください。Helperを単に削除しても自動では戻りません。
`Auto`は全physical groupのVideo Decodeをプロセス専用Diskへ、小さなAudio Decodeを上限付きRAMへ保存します。`Off`はnative Decodeを使い、Helper Cacheを破棄します。Cache障害時はnative Decodeへ戻りますが、native DecodeのOOMやQueue中断は再試行しません。Python再起動後にCacheファイルを再利用しません。移行時は同じ`H3DecodeCacheHelper` IDを提供する旧単独addonと内蔵版を同時登録しないでください。内蔵版の導入確認後に旧addonを無効化または撤去します。
736×416(0.306MP)、3×5秒、Euler/simple 6 steps、Turbo FL2V v1.2、ウォーム、Video 3/3 HITの1条件では、Full Run Totalの**機構推定中心値は約47秒/約24.5%短縮**でした。直接A/Bの1ペアは50.887秒/26.46%、Samplingを伴わないComplete/Resumeの2ペアは平均46.309秒/79.826%短縮でした。これは限定条件の観測値で、反復された一般平均や無条件の20%以上高速化保証ではありません。初回MISSは遅くなる場合があり、新規Samplingは高速化せず、latentまたはVAE識別情報が変わればMISSになります。
MiniMax H3を複数チャンクで連続生成し、直前チャンク末尾の**映像latent / 音声latentを直接**次チャンクへ継承するComfyUIカスタムノードです。チャンク間でVideo/Audio VAEのDecode→Encodeは行いません。
## インストール・更新
新規導入:
```bash
cd ComfyUI/custom_nodes
git clone https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum.git
```
既存のGit導入を更新:
```bash
cd ComfyUI/custom_nodes/ComfyUI-H3-Continuum
git pull --ff-only origin main
```
ComfyUIを再起動してください。Managerから導入した場合はManagerのUpdateを使い、同じノードを重複配置しないでください。
旧V3.8.0を別フォルダへ導入する場合:
```bash
cd ComfyUI/custom_nodes
git clone --branch v3.8.0 --single-branch https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum.git ComfyUI-H3-Continuum-v3.8.0
```
Continuumのcheckoutは同時に1つだけ有効にしてください。ComfyUI Managerの**Update**は`main`を更新するため、過去tagの選択には使いません。
V3.8X2 Workflow:[JSON](../examples/workflows/MiniMax_H3_Continuum_V38X2.json)/[V3.8X2・V3.9両JSONを含むZIP](../examples/workflows/MiniMax_H3_Continuum_V38X2.zip)。V3.8X2 JSONはSpectrum対応の1本で、保存状態ではSpectrumとTurbo LoRAはどちらもOFFです。[LightX2V Turbo](https://github.com/ModelTC/Minimax-H3-Turbo)にも切り替えられます。完全なgraphを開くにはSpectrum・rgthree・KJNodesが必要です。ComfyUI-Easy-Useは不要です。
## プロンプト・スキルのダウンロード
- [V3.9 Reference Images用プロンプトスキル本文](../write-h3-v39-reference-prompts/SKILL.md)/[ZIP](../write-h3-v39-reference-prompts.zip):固定の`@R1`~`@R9`とチャンク別の画像割り当てに対応します。V3.9のReference Imagesノード用で、下記の汎用Continuumスキルとは別です。
- [LLM用システムプロンプトZIP](../H3-Continuum-LLM-System-Prompt-v1.zip):Continuum向けプロンプト作成のシステム指示と詳細資料。
- [Continuum専用プロンプトスキルZIP](../H3-Continuum-Skill-v1.zip):Codexなどで使う、チャンク構成に対応した汎用プロンプト作成スキル。
- [Continuum Dance DirectorスキルZIP](../minimax-h3-continuum-dance-director.zip):主役のダンサー1人を対象に、チャンク間の身体動作とカメラの連続性を考慮した長尺ダンス用プロンプトを作成するスキル。
いずれも任意のプロンプト作成支援資料で、ComfyUIのカスタムノードではありません。ZIPを展開し、同梱の案内に従って利用してください。Samplerを変更するものではなく、生成動作や画質を保証するものでもありません。
## V3.8の公開範囲
H3 Continuumは、長尺映像を最初からやり直さずに生成・確認・部分再生成・再開するProduction Samplerです。V3.8は`Main / Production`と`Advanced`の2層に整理しました。
以下の9ノードに`H3 Continuum Video Adapter`を加えた計10 IDがV3.8X2の登録面です。旧Audio/Video Loaderは互換IDとして残ります。V3.9はさらにSampler V3.9とReference Images V3.9を追加した計12 IDです。
- `H3 Continuum Sampler V3.8`
- `H3 Continuum Finalize`
- `H3 Continuum Load Image`
- `H3 Continuum Load Audio`
- `H3 Continuum Load Video`
- `H3 Continuum Second Pass`
- `H3 Continuum Reference Audios`
- `H3 Continuum Reference Images`
- `Decode Cache Helper`
Main Sampler内の`Show Advanced Settings`/`Hide Advanced Settings`で詳細設定を開閉します。旧node property `H3 Continuum View`を含む保存Workflowは対応する開閉状態へ移行し、旧propertyを削除します。これは表示状態だけの変更です。Python input、widget順、保存済みwidget値、backend引数、Sampling identity、Run Storage identityは変更しません。Frontend extensionが読み込まれない場合も、Python定義の全widgetが表示された状態で実行できます。
V3.8X2公式Workflowの標準経路は`Sampler V3.8 -> Decode Cache Helper -> Finalize -> Create Video -> Save Video`です。HelperはMISSしたentryをCore Video/Audio Decodeへ委譲します。Core Decodeへ直接つなぐ手動構成も使用できますが、Decode結果のcacheはありません。Finalizeの`IMAGE`/`AUDIO`出力をCoreの`Create Video`で`VIDEO`へまとめてから保存します。Save、Upscalerは外付けとし、外部latent processorやupscalerはSecond Passをbridgeとして接続します。外部processorが変更できるのはVideo LATENTの空間解像度だけで、physical group、B/C/T、finite値、First Pass Audio、元のAssembly Planを維持する必要があります。Finalizeは公開`IMAGE`/`AUDIO`型を受け取り、特定decoder classへ依存しません。SageAttention、Sol-Attn、Spectrumは外部MODEL wrapperのままです。詳細は[V3.8 Open Integration Contract](V38_OPEN_INTEGRATION_CONTRACT.md)を参照してください。旧Hi-Res FixはV3.8X2標準Workflowに含めません。
### Reference Audio 1/2/3
`H3 Continuum Reference Audios`は、最大3本の単独Reference Audioを1本の`Audio References (Optional)` socketへまとめます。入力は間を空けずに接続し、Promptでは接続順どおりに`<Audio 1>`、`<Audio 2>`、`<Audio 3>`を使用します。共通のCore Audio VAEで各Audioを個別encodeします。生成後の最終Audioを置換せず、Driving Audio契約も変更しません。既存Workflowは従来の単数`Reference Audio (Optional)`をそのまま使用できますが、単数経路とbundle経路は同時接続しないでください。
> **対応範囲:** 元のV3.8公開面は7 ID、V3.8X2は互換Loaderを含め10 ID、V3.9は12 IDです。Finalizeの`H3ContinuumAssembleSeamV35`、Second Passの`H3ContinuumSecondPassV35`などは旧IDを維持しますが、すべての旧Workflowに互換性があるという意味ではありません。未対応のnode IDには対応するhistorical Release/tagを使用してください。詳細は[V3.8 Release/Migration Policy](V38_RELEASE_AND_MIGRATION.md)を参照してください。
複雑な16GB GPU受入Gateでは約`15.5~15.6 GiB`を使用しました。GPU、driver、backend、model精度、解像度、接続ノードで変動するため、すべての16GB GPUでの動作保証ではありません。
Reference Imageは最大9枚です。Samplerの直入力1~3を維持し、追加4~9は`H3 Continuum Reference Images`のoptional IMAGE入力へ接続して、出力をSamplerの`Reference Images (Optional)`へ渡します。補助ノードは画像を束ねるだけで、Resize・hash・VAE Encodeは既存Sampler経路で実行します。未接続を除き1→9の順に使い、Picture番号はFirst/Last Imageの後に連番で割り当てます。空のbundleは参照なしとなり、旧0~3枚の生成・再利用契約は変わりません。
以前の4・5直入力を含む保存Workflowは、グラフ読み込み後に元の画像接続をbundleへ移します。新接続を検証した後だけ旧接続口を削除し、接続元が不明・移行先が使用中の場合は警告して元の線を残します。旧API入力も維持しますが、同じ番号の画像を旧直入力とbundleの両方から指定すると入力競合として明示します。既存の3枚loaderテンプレートは上書きせず、そのまま利用できます。
### サイズの決め方を直接選択
| Preset | 目安 |
|---|---|
| `Draft` | 約0.30 MP |
| `Balanced` | 約0.60 MP |
| `Native 768` | 短辺768を目標。極端なAspectでは長辺1344 capを優先 |
| `Custom MP` | Megapixelを直接指定。Native 768のcapとは別 |
Mainの`Size Source`は次の2択です。
- `First Image`:接続したFirst Imageの縦横比を維持し、Presetの目標面積から最終Width/Heightを算出します。
- `Manual`:Width/Heightを32 pixel単位で直接指定します。この場合、Presetは適用しません。
MainのWidth/Heightは`Manual`のときだけ表示・編集できます。`First Image`ではResolutionを表示し、画像の縦横比からサイズを決めます。有効なFirst ImageがSamplerへ届かない場合は、保持しているManual Width/Heightへフォールバックし、その寸法をStatusへ表示します。手動寸法を確認・変更する場合は`Size Source = Manual`を選びます。Reference ImageやVideo Guideからサイズを暗黙取得することはありません。
旧V3.8の`Auto / Landscape / Portrait / Square`はWorkflow/API互換用として内部で受け付けます。FrontendはAutoをFirst Imageへ、3つの固定Aspectを同じ解決結果のManual Width/Heightへ移行します。
### 10秒を5秒ずつ確認しながら作る
ComfyUI上端の実行ボタンは、frontendにより`Queue`、`Run`、または`実行する`と表示されます。この手順内の`Queue`は、その実行ボタンを指します。
ここでは、画面に表示される名前だけを使って手順を説明します。
#### 最初に知っておくこと
`Chunks = 2`、`Seconds per Chunk = 5`は、完成目標が合計10秒という意味です。ただし、`Run = Review Each Chunk`では10秒を一度に生成しません。最初のQueueでは、確認用に**Chunk 1の5秒だけ**を生成します。これは正常動作です。
Chunk 1を確認した後に操作を選び、もう一度QueueするとChunk 2へ進みます。`Chunks = 1`では最初の5秒が最後なので、次へ進む操作はありません。
#### 手順1:生成前の設定
`H3 Continuum Sampler V3.8`で、上から次の順に設定します。
1. `Control After Generate`を`fixed`にします。`randomize`、`increment`、`decrement`では続きを再利用できません。
2. `Chunks`を`2`にします。
3. `Seconds per Chunk`を`5`にします。
4. `Total Length`が`10 seconds`になったことを確認します。`Total Length`は確認表示で、直接変更する項目ではありません。
5. `Run`を`Review Each Chunk`にします。
6. `Progress`を`On — Resume and Takes available`にします。
7. 緑色の`Ready to Queue`欄に、`2 × 5s = 10 seconds`と`Review each chunk`が表示されていることを確認します。`Set Control After Generate to fixed`と表示された場合は、Queueする前に手順1を直します。
#### 手順2:最初の5秒を作る
1. ComfyUIの`Queue`を押します。
2. Chunk 1だけが生成されます。出力動画が5秒でも正常です。
3. Queueの処理が完了すると、Samplerの表示が自動的に`Chunk 1 is ready for review`へ切り替わります。
#### 手順3:Chunk 1の動画を確認する
保存された5秒の動画を再生し、そのまま採用するか、作り直すかを決めます。この時点ではまだ合計10秒の動画は完成していません。
#### 手順4:次の操作を1つ選ぶ
| 画面上のボタン | 次のQueueで行うこと |
|---|---|
| `Use it and continue` | 現在の5秒を採用し、次のChunkを1つ生成します。 |
| `Try this chunk again` | 現在のChunkだけを別のTakeとして作り直します。 |
| `Use it and finish the rest` | 現在までを採用し、残りのChunkをまとめて生成します。 |
ボタンを選んだだけでは生成は始まりません。選択したボタンに`✓`が付いたことを確認してから、ComfyUIの`Queue`を押します。
`Try this chunk again`はContinuum側で別Takeを作るため、作り直しでも`Control After Generate`は`fixed`のままにしてください。
#### 手順5:2本目の5秒を作り、10秒を完成させる
1. `Use it and continue`を選びます。
2. ComfyUIの`Queue`をもう一度押します。
3. 採用済みのChunk 1を再利用し、Chunk 2を生成します。
4. Samplerから後ろのDecode、`H3 Continuum Finalize`、Save Videoまで通常どおり実行されると、合計10秒の完成動画が出力されます。
設定を確認し直したい場合は`Back to Settings`を押します。未処理のReviewへ戻る場合は`Return to Review`を押します。どちらも生成を自動開始せず、選択済みのTakeや保存済みChunkも削除しません。
#### 迷ったときの確認
- 最初のQueueで5秒だけ出た:`Review Each Chunk`の正常動作です。
- 次へ進むボタンが出ない:Queueが完了しており、`Progress = On — Resume and Takes available`、`Chunks = 2`以上であることを確認してください。正常なら完了後に自動表示されます。
- ボタンを押しても生成されない:操作を選んだ後、ComfyUIの`Queue`をもう一度押してください。
- `Set Control After Generate to fixed`と表示された:`Control After Generate`を`fixed`にしてください。別SeedのChunk 1が増えることを防ぐ安全チェックです。
- 1回で10秒すべて作りたい:`Run`を`Generate Full Video`にします。
Long Terminal Mergeのpairは分割せず、1つのReview単位として扱います。
Driving AudioはReview/Smart Regenerateと併用できます。Continuum Image/Audio/Video Loaderは任意入力向けのnative bypassに対応します。
> **V3.8 Reviewの制限:** Review途中のpartial sequenceでは、Second Pass/`refine_context`を正式対応範囲に含めません。Reviewを完了してsequenceを確定してからSecond Passを実行してください。V3.9では、上記のとおり検証済みの出力区間条件を継承できます。
### Take・Branch・安全な継続
Run Storageを有効にすると、Production表示に**Render History / Takes**が表示されます。
Takeを選択してもcanonical結果は変わりません。**Use This Take**でそのrevisionを
canonicalにし、**Continue From Here**で選択したTakeまでを保持して後続groupだけを
再生成します。どちらの操作も自動Queueは行いません。