ComfyUI-H3-Continuum 3.9.1 / V3.9 stability validation - 2026-10-04 Current package metadata: 3.9.1. Public registration: twelve nodes. The V3.9 Sampler and Reference Images helper retain their public IDs and official workflows; V3.8X2 remains a separate workflow. LoRA Plan / V4.0 is not included. Run Storage v3 and State/Session formats are unchanged. Repair scope - Preserve audio-only/mixed Second Pass keyframes; spatially adapt video only. - Match partial/middle-chunk Review conditioning and References to the actual output physical groups, retaining Terminal Merge groups as atomic units. - Use Sampling Contract v6 / Graph Contract v4 for Take compatibility, including sampler closures and MODEL CFG/wrapper settings. - Crop Review Driving Audio once from the original physical-group time range. - Prefer Core audio resampling. Direct weight mutation inside the same VAE object requires explicit Reference Encode Cache clearing or VAE reload. Recorded validation (separate environments; do not add these counts) - Full CPU with real ComfyUI Core: 1,696 passed / 1 skipped / 0 failed. CUDA initialization attempts: zero. - Release preparation with real Core: 66 passed / 0 skipped / 0 failed. - Standalone dependency check: 48 passed / 18 skipped / 0 failed. - Release document/manifest checks: 9 passed / 0 failed. - GitHub Actions at release commit 94eaf70d674a58244e1cb139f021290d63fb2e6c: 1,676 passed / 21 skipped / 0 failed. Eighteen cases require real Core and are explicitly skipped without it; three retain existing environment or fixture limits. This is not 1,696 real-Core cases executed in CI. - Offline native PackedLayout, Fixed 3x5-second planning and twelve-node registration passed. - Targeted GPU partial-prefix, single-middle-chunk and complete-sequence Review Second Pass checks passed at 512x512 / 24 fps, including conditioning inheritance, source-PCM alignment and saved Take/raw preservation. - Integrated generation and Take-reuse GPU checks passed for the repaired runtime. These are targeted results, not universal quality/accelerator gates. Current manifests declare 429 development files and 144 Registry payload files. Hash integrity does not establish semantic freshness of every document. The official V3.9 workflow retains 32 nodes / 39 links, the original JSON bytes and matching workflow ZIP. Models/media are not included. Publication and remaining gates - GitHub Release v3.9.1 is published. Its tag remains fixed at the release commit above; later documentation corrections are follow-up main updates. - Registry publication is separate and has not been performed for this patch. - Browser save/reload acceptance and full acceptance of the unchanged default workflow remain separate gates; do not infer them from CPU or targeted GPU results. Record the tested workflow and runtime when completing them. - Optional latent-upscale doubled outlines remain separate in Issue #27. Updating: restart ComfyUI and refresh the browser. Old v5 Takes/history remain readable and are not deleted or rewritten, but cannot supply prefixes to new v6 generation. Start a new Full Video or Start again from Chunk 1. Continue using the V3.9 nodes. See README.md, README_JA.md and docs/VALIDATION_JA.md for current instructions. ====================================================================== Historical V3.9.0 main-source record - 2026-09-29 The original record below is retained as history. Its metadata, test counts, manifest counts and publication status do not describe package 3.9.1. ====================================================================== ComfyUI-H3-Continuum 3.9.0 / V3.9 main-source validation — 2026-09-29 Status: MAIN SOURCE PUBLICATION. The ComfyUI_W worktree has 12 public node IDs: the existing V3.8 and V3.9 Samplers plus the new V3.9 Reference Images helper. Package metadata is 3.9.0. A GitHub main commit is not a new GitHub Release or Registry publish; earlier Releases are retained. The user-selected official V3.9 JSON is copied byte-for-byte and has 32 nodes/39 links. Its matching one-JSON ZIP and gallery copy preserve that graph. The standard V38X2 ZIP now includes both unchanged V38X2 and V3.9 JSONs for one-download setup; the V38X2 JSON itself is unchanged. V3.9 retains the V38X2-based model/LoRA, Decode Cache, Finalize and Save chain; Image loaders 1–6 are wired to the new helper but bypassed by default, while 7–9 require user-added loaders. Its Sequence Prompt is blank. Its saved size controls are First Image + Draft 0.30 MP, with the First Image loader enabled; users must select their own image before Queue. For Reference-only, turn it off and choose Manual. Saved 480x640 applies only to Manual mode. Other saved controls include two 10-second chunks, Full Run mode, Spectrum/Turbo LoRA off, and a separate Project ID. V3.9 connects fixed Reference Images 1–9 to one helper, then connects its one typed output to the V3.9 Sampler. The helper's in-node table selects All chunks or per-chunk assignments. The backend retains sparse fixed IDs, uses the existing effective plan/Run Storage contracts, and rejects incompatible payloads. The V3.8 Sampler and official V38X2 workflow JSON/ZIP are unchanged. Older V3.9 prototype workflows need manual rewiring; saved Runs/Takes are not automatically migrated. The latest recorded Windows CPU run is 1536 passed, 1 skipped, 0 failed. A separate Linux/Python 3.13.5 CPU audit reported 1534 passed, 3 skipped, 0 failed; that external count was not rerun during this documentation edit. The selected workflow JSON has 32/39 nodes/links. The development Manifest contains 406 entries and the Registry Manifest contains 138 entries. Both were resynchronized and hash-checked after this documentation reconciliation. The earlier RR-R6B browser result predates this helper UI, and configured Reference-only GPU/API evidence does not establish browser save/reload or GPU acceptance of the current default official template. That acceptance remains HOLD. This source publication makes no generation-code changes relative to the validated local V3.9 candidate and does not create a tag, GitHub Release, Registry package, or new GPU run. Historical V3.8X2 / package 3.8.3 validation record follows. Its test counts, node count and workflow SHA describe that earlier preparation, not the current V3.9 candidate: ComfyUI-H3-Continuum V3.8X2 / package 3.8.3 — package validation record Status: local commit/push preparation PASS. No commit, push, tag, GitHub Release, or Registry publication is authorized or performed by this record. Public surface: exactly nine nodes - H3 Continuum Sampler V3.8 - H3 Continuum Reference Audios - H3 Continuum Finalize - H3 Continuum Load Image - H3 Continuum Load Audio - H3 Continuum Load Video - H3 Continuum Second Pass - H3 Continuum Reference Images - Decode Cache Helper Official V3.8X2 workflows - MiniMax_H3_Continuum_V38X2.json / .zip - MiniMax_H3_Continuum_V38X2+Decode_Cache_Helper.json / .zip - Both names contain the same graph. The current standard V38X2 ZIP also includes V3.9; the helper-explicit ZIP remains single-JSON. The historical V38X2 JSON SHA-256 is 6c445d62979632bc1a801b1c28dbd639c23b51cc156ebae18837baaa757e592b. Reference Image 1–9 resource gate - Tested system: NVIDIA GeForce RTX 5060 Ti 16 GB / RAM 64 GB. - Tested run: 512 x 608, 3 x 5 seconds, Full Run, Spectrum Off, LoRA Off, Sage Attention, Balanced 22-frame continuity. - Nine test references were reduced to about 0.30 MP each. - Sampling used about 15.0–15.5 GiB of the available 16,311 MiB VRAM. - This is a high-resource configuration with little 16 GB headroom. Pre-resize source-image copies to about 0.30 MP before loading all nine. Larger inputs, output sizes, wrappers, or concurrent workloads may require more VRAM/RAM. Decode Cache Helper boundary - It can reuse Decode only when latent contents, shape, dtype, decode metadata, VAE identity, and a process-local cache entry all match. - It does not accelerate Sampling. A first MISS can be slower. - The measured 20%+ reductions are configuration-specific observations, not an unconditional speed guarantee. Preparation verification - Full CPU suite: 1,367 passed / 1 skipped / 0 failed. - Frontend Review Queue: 46/46 PASS. - Workflow JSON structure: 32 nodes / 38 links / zero broken saved links. - Workflow aliases: byte-identical JSON; ZIP payload parity required. - Development and Registry manifests: regenerated after documentation and workflow bytes were finalized; final hash verification is part of the commit-ready Gate. Registry package policy - Include production runtime modules and only the explicitly re-included tools, documents, and four V3.8X2 workflow files from .comfyignore. - Exclude tests, internal AGENTS/PROJECT_STATE/WORKLOG files, development MANIFEST.sha256, historical workflow files, caches, and unrelated ZIPs from the Registry payload. - MANIFEST.sha256 covers the prepared source tree except itself and the three internal project-state files. REGISTRY_MANIFEST.sha256 covers the actual Registry payload except itself.