Release v3.6.0 masked AV continuation
This commit is contained in:
+102
@@ -1,5 +1,107 @@
|
||||
# Work Log
|
||||
|
||||
## 2026-08-27 - V3.6.0 public release implementation and GPU smoke
|
||||
|
||||
- Added `H3 Continuum Sampler V3.6` as a new public Node ID while preserving V3.5.3/V3.4 nodes and widget sequences. Public backend choices are only `Standard` and `Compatibility`: Standard maps to Masked AV or Masked Video according to Audio Continuity, while Compatibility maps to the original Reference Context path. Internal transport names remain private.
|
||||
- Updated package/version metadata, CHANGELOG, package validation/info, README English/Japanese, compact frontend handling, runtime verifier registration, and a copied V3.6 recommended workflow. The V3.5 workflow remains unchanged. Added public-node/backend/schema/workflow regressions.
|
||||
- Synchronized only `nodes.py`, `version.py`, `metadata.ini`, `v3/driving_nodes.py`, and `web/project_id.js` to ComfyUI_WAN after snapshot `pre-v360-public-runtime-wan-20260827_172531`; all five hashes matched. Restarted the same Sage + `--disable-pinned-memory` backend and confirmed V3.6 registration with `Standard` default.
|
||||
- Normal API GPU Smoke using the public node completed in `65.21s`: 2x5-second I2VA + Reference Image + Reference Audio, 384x384, four steps, Standard backend, Audio Seam Auto. Output `V36_Release_Smoke_Reference_AV_Standard_00001_.mp4` is 240f / 24fps / 10.000s with 32 kHz stereo 10.000s audio; the report confirms Masked AV on continuation chunk 2. Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v360-release-smoke\20260827_172835`.
|
||||
- Focused release validation is `81 passed`; final full source pytest is `527 passed`; V3.6.0 runtime verifier, JavaScript syntax, changed-file `py_compile`, the regenerated 178-entry Manifest, README link validation, source/runtime SHA-256 equality for 12 Production files, `git diff --check`, and empty API queue all pass. Commit/push/release are the remaining publication actions; the backend remains running.
|
||||
|
||||
## 2026-08-27 - V3.6 PIG-5 subjective audio acceptance
|
||||
|
||||
- User completed subjective listening of `V36_PIG3_Reference_Image_Audio_Masked_AV_00001_.mp4` and reported **Audio PASS**. No click, dropout, unnatural stereo positioning, or other audible boundary defect was reported.
|
||||
- PIG-0 through PIG-5 and V3.6 public-promotion eligibility are now formally PASS. This records the gate result only: published V3.5.3, public UI/defaults, version metadata, node schemas, runtime files, and workflows were not changed in this step.
|
||||
- Pre-change documentation snapshot: `pre-pig5-audio-pass-20260827_171054`. Commit/push were not performed.
|
||||
|
||||
## 2026-08-27 - V3.6 PIG-5 policy freeze
|
||||
|
||||
- User approved the final Production policy: Masked AV is the standard V3.6 continuation backend; Reference Context remains an Advanced compatibility fallback; raw transport names stay internal; Run Storage identity stays transport-scoped; Terminal Merge regeneration remains atomic per physical group; and Audio Seam Auto continues as an independent evaluation path.
|
||||
- Public promotion remains HOLD only for subjective listening of the accepted PIG-3 output. A clean result for clicks, dropouts, stereo-positioning anomalies, and other audible boundary defects is sufficient to promote V3.6 without reopening PIG-0 through PIG-4. No UI, runtime, node schema, Sampling, Run Storage, or public default was changed in this policy-recording step.
|
||||
- Pre-change documentation snapshot: `pre-pig5-policy-freeze-20260827_170808`. Commit/push were not performed.
|
||||
|
||||
## 2026-08-27 - V3.6 Production Integration Gate PIG-0 through PIG-4
|
||||
|
||||
- Preserved published/default V3.5.3, every accumulated V3.6 change, and all unrelated working-tree edits. Accepted exact snapshots are `pre-v36-production-integration-pig0-accepted-20260827_161517` and `pre-v36-production-integration-pig2-diagnostic-accepted-20260827_162943`; the standard full snapshot stopped only on the known inaccessible `.pytest_cache` before edits.
|
||||
- Added nondefault `continuation_transport` scoping to Run Storage identity, Sampling Contract execution semantics, and generated Session settings while leaving the default `reference_context_v1` revision identity bit-for-bit compatible. Added normal/Terminal Reference Image + Reference Audio CPU regressions and transport/session/terminal identity tests.
|
||||
- PIG-2 real-cache testing exposed that Run Storage's temporary resume Session omitted stored `flf_execution`, causing all preserved chunks to be discarded. `run_storage.py` now restores execution semantics into the resume Session, and `v2/sequence.py` synchronizes the reuse count after atomic Terminal-pair pruning. The accepted rerun produced `0/3`, `3/0`, `1/2`, and `1/2` reused/generated counts for Save, Auto Resume, Regenerate From Chunk 2, and Chunk 3; the two regeneration requests each made exactly one Terminal physical Sampling call and never split logical `[2,3]`.
|
||||
- PIG-3 completed 2x5-second I2VA + Reference Image + Reference Audio through Masked AV at 384x384 / 4 steps. Every physical group retained both Reference kinds, protected Video/Audio prefixes were bit-exact after finalization, and the output is finite 240f / 24fps / 10.000s with 10.000s stereo audio. Existing Audio Seam Auto produced a benign retained-boundary jump (`0.004261`, `0.0876x` local p99); subjective listening remains a separate user check.
|
||||
- Evidence roots are `D:\Codex\_test_results\ComfyUI-H3-Continuum\v36-production-integration\v36_pig2_masked_20260827_164303` and `...\pig3-reference-masked`. Production integration is PASS; public/default promotion remains HOLD pending explicit UI/default-policy approval and subjective Audio Seam listening. No public UI was added, commit/push were not performed, and the backend remains running.
|
||||
- Final validation: PIG-focused `61 passed`; canonical full pytest `521 passed`; changed-file `py_compile`, V3.5.3 WAN runtime verifier, three source/runtime SHA-256 comparisons, `git diff --check`, and empty API queue PASS. Only the known `pynvml` FutureWarning remains.
|
||||
|
||||
## 2026-08-27 - V3.6-R2 shortened FL2VA Long Terminal Merge final GPU gate
|
||||
|
||||
- Preserved published/default V3.5.3 and every accumulated uncommitted V3.6 change. Snapshot `pre-v36-r2-short-gate-20260827_000840` contains the exact runner/docs changed for this gate. Added only `--phase r2-short` scheduling to the private gate runner: Seed 607 full Reference/Masked comparison with cache-only Seam OFF, followed by Seed 608 Masked-only acceptance.
|
||||
- Combined the existing complete Seed 606 Gate A with the shortened runs. All accepted executions preserve 3 logical chunks, 2 physical groups, Group 2 logical `[2,3]`, `paired_timeline_v1`, terminal seed/trim, and `260f / 77T Video / 433T Audio`. Every main MP4 is finite 640x640, 360 frames, 24 fps, and exactly 15 seconds with 15-second audio.
|
||||
- Seed 607 Reference/Masked Group 1 Video and Audio SHA-256 values match. Seeds 606–608 Masked finalization restores only protected 7T Video and 37T Audio prefixes to bit-exact equality; raw Core drift remains limited to one-float-scale deltas and generated regions remain sampler-owned.
|
||||
- Terminal PackedLayout is `35,865 -> 32,991` rows, a `2,874`-row / `8.01%` reduction. Synchronized Group 2 Sampling pairs are `574.134 -> 529.933s` and `560.782 -> 544.413s`; both favor Masked. Paired percentage median is `5.31%`; method medians are `567.458s` Reference and `544.413s` Masked (`4.06%` lower). Seed 608 Masked completed in `549.704s` Group 2 / `801.006s` API wall time.
|
||||
- Five-run monitored peak RSS median is `9.674 GiB`; maximum observed device use is `15.646 GiB`. Terminal CUDA allocator peak is approximately `4.06 GiB` Reference versus `3.80 GiB` Masked. These whole-process/device values are environment- and order-sensitive; no broad memory guarantee is inferred.
|
||||
- Inspected Seed 606–608 boundary sheets at nominal 5/10/15-second positions and actual 124f/243f retained boundaries. No obvious flash, collapse, direction reversal, or temporal discontinuity was found. Audio analysis found finite 15-second streams and no >=200 ms -50 dB dropout. Audio Seam Auto materially improved Seed 606 Masked, while Seed 607's 124f instantaneous jump increased despite no dropout or structural failure; the report therefore records mixed numerical seam behavior rather than claiming universal improvement.
|
||||
- Final checks: canonical full pytest `516 passed` with only the known `pynvml` warning; the first attempt's 98 setup errors were solely the inherited inaccessible Windows pytest temp root and disappeared with a new writable basetemp. WAN runtime verifier, source/runtime SHA-256 for `constants.py`, `v2/nodes.py`, `v2/prompts.py`, `v2/sequence.py`, `v3/nodes.py`, `v3/masked_continuation.py`, and the diagnostic node, `git diff --check`, and empty queue pass.
|
||||
- Gate result: **PASS**. Production promotion: **HOLD** pending Reference variants, Run Storage/resume identity, public UI/default decisions, and broader audible Audio Seam acceptance. Evidence roots are `D:\Codex\_test_results\ComfyUI-H3-Continuum\v36-fl2va-combined-20260826\gate-a` and `...\r2-short`. Backend PID 15888 remains running per request. Commit/push were not performed.
|
||||
|
||||
## 2026-08-27 - V3.6 shared chunk-duration maximum 30 seconds
|
||||
|
||||
- Preserved all uncommitted V3.6 R1/R2/FL2VA work and changed only the shared duration range, Prompt Plan validation, documentation, and focused regressions. No per-chunk variable-duration feature, Sampling, Conditioning, `17k+5` temporal math, Assembly, Seam, Run Storage, or saved-workflow socket contract changed.
|
||||
- Added shared constants for default `5.0`, minimum `4.0`, maximum `30.0`, step `0.1`, and one tooltip. Applied them to all V2/V3 Production-compatible sampler schemas and Prompt Plan Preview. Values above 15 seconds are supported with a high-resolution VRAM/runtime warning; `5–15s` remains the recommended/validated range.
|
||||
- Prompt Plan creation and connected-plan validation now accept `30.0`, reject `30.1`, and correctly map a two-chunk `[0-30s]` / `[30-60s]` Timeline. Temporal regression covers every tenth from `4.0` through `30.0` for 5/22/39-frame contexts and requires native `17k+5` output shapes.
|
||||
- The separate legacy V1 `H3ContinuumJoin.extend_seconds` maximum remains 15 seconds by design; the audit found no remaining 15-second cap on `chunk_seconds`.
|
||||
- Source snapshot `pre-v36-chunk-duration-30-source-20260826_234716` verifies all 209 Git-visible files. Runtime snapshot `pre-v36-chunk-duration-30-wan-runtime-20260826_235423` verifies the four synchronized files. Source/runtime pre-change differences in `constants.py` and `v2/prompts.py` were EOL-only.
|
||||
- Validation: focused inherited-schema/prompt/temporal suite `89 passed`; final full pytest `516 passed`; changed-file `py_compile`, `git diff --check`, V3.5.3 runtime verifier, source/WAN SHA-256, live API schemas for seven V2/V3 node IDs, and empty queue PASS. The restarted WAN backend uses Sage Attention plus `--disable-pinned-memory` and is intentionally left running. Commit/push were not performed.
|
||||
- Per the reduced GPU scope, submitted one T2VA `1x30s`, `384x384`, one-step validation prompt. API returned prompt ID `9e665f85-b32f-4b17-9734-219fb284e087` with `node_errors={}`; H3 model initialization/Sampling began without immediate duration/shape/Prompt Plan/OOM error. Interrupted intentionally after `31.87s`; no Decode/Save or 30-second output-quality claim was made. Queue returned empty.
|
||||
- Cancelled the FL2VA joint multi-seed continuation gate at user request. Its cache warm-up completed, but the first measured condition was interrupted before execution and is excluded from results.
|
||||
|
||||
## 2026-08-26 - V3.6-R2-2 Balanced 22f / 37T Joint AV GPU gate
|
||||
|
||||
- Preserved the public/default V3.5.3 `reference_context_v1` route and all accumulated R1/R2 work. Added only private `masked_av_prefix_22_v1` handling and diagnostic tooling; no public node/schema, V3.4 behavior, `v2/sampling.py`, seed, Plan, trim, Decode, Assembly, Seam, Run Storage, or saved-workflow contract changed.
|
||||
- Pre-change exact-file snapshots are `pre-v36-r2-2-balanced-av-source-files-20260826_214033` (9 files, zero mismatch) and `pre-v36-r2-2-balanced-av-wan-runtime-files-20260826_214033` (3 files, zero mismatch) under `D:\Codex\_snapshots\ComfyUI-H3-Continuum`. The initial runtime snapshot copy omitted subdirectories and was corrected before any edit; only the verified snapshot is accepted.
|
||||
- CPU/static validation established the Balanced target `141f / 42T / 235T`, Video prefix `22f / 7T`, Audio prefix `37T`, trim `22f`, net `119f`, and the `8.333ms` Video/Audio mask-boundary asymmetry. Focused result is `24 passed`; pre-GPU full result is `510 passed` with only known warnings.
|
||||
- Synchronized only `v2/sequence.py`, `v3/masked_continuation.py`, and private diagnostic `nodes.py` to ComfyUI_WAN with SHA-256 equality. Restarted ComfyUI v0.34.0 with Sage Attention and `--disable-pinned-memory`; node registration and empty queue passed.
|
||||
- One-seed A/B used T2VA 2x5, 640x640, Balanced 22, Audio Continuity ON, fixed seed `238985343135605`, Prompt/CLIP cache HIT, CUDA synchronization, and Exact Duration ON. Reference packed layout was `20,205` rows; Masked AV was `17,331`, eliminating `2,874` reference rows (`14.22%`). Continuation Sampling was `226.373s -> 206.437s` (`8.81%` lower) and API wall time `435.907s -> 420.752s` (`3.48%` lower). One-seed timing is not generalized.
|
||||
- Masked source/target Video 7T and Audio 37T prefixes match bit-exactly after scoped final restore. Core raw prefix deltas were only `2.384185791015625e-07` Video and `5.960464477539063e-08` Audio. Both outputs are finite `640x640 / 240f / 24fps / 10.000s`; boundary contact sheets show no obvious visual regression.
|
||||
- Seam-OFF decoded audio is not intrinsically accepted: Reference jump `0.082725`, Masked jump `0.268163`. Existing Audio Seam Auto reassembled the cached Masked latents in `7.10s`, did not rerun Sampling, and reduced the jump to `0.029944` (`88.83%` lower and below local p99). No new Audio Seam code was added.
|
||||
- R2-2 status: GPU structure/performance gate **PASS with existing Audio Seam Auto**, raw Seam-OFF audio documented, Production **HOLD**. Final full pytest is `510 passed`; changed-file `py_compile`, `git diff --check`, source/runtime SHA-256, V3.5.3 runtime verifier, and empty queue pass. Evidence is under `D:\Codex\_test_results\ComfyUI-H3-Continuum\v36-r2-gate2-20260826`. Backend remains running; commit/push were not performed.
|
||||
|
||||
## 2026-08-26 - V3.6-R2-1 39f / 65T Joint AV GPU structure gate
|
||||
|
||||
- Synchronized only `v2/sequence.py`, `v3/masked_continuation.py`, and the private diagnostic `nodes.py` to ComfyUI_WAN after the accepted R2-0 runtime snapshot. All three source/runtime SHA-256 hashes match. Restarted the same ComfyUI v0.34.0 backend with Sage Attention and `--disable-pinned-memory`; the backend is intentionally left running.
|
||||
- Added reusable `tools/v36_r2_gpu_gate_runner.py`. It removes the I2VA First Image, fixes T2VA 2x5 / 640x640 / Strong 39 / Audio Continuity ON / one seed, excludes optional conditioning and Run Storage, warms Prompt/CLIP once, and measures Reference then Masked with CUDA synchronization. A validation-only stop found Core reports the 39f AV Reference as combined kind `video_audio`; the completed Reference generation remained valid and Masked resumed without repeating it.
|
||||
- Reference continuation used `24,317` PackedLayout rows (`4,800 ref_img`, `130 ref_audio`); Masked used `19,387` with no Reference rows. The exact reduction is `4,930` rows / `20.27%`. Synchronized Chunk 2 time was `332.460 -> 241.888s` (`27.24%` lower); API wall was `556.058 -> 465.925s` (`16.21%` lower). This is one-seed structural evidence, not a universal performance claim.
|
||||
- GPU evidence confirmed source-tail/target-prefix equality, Video mask `0:12` protected and `12:47` generated, Audio mask `0:65` protected and `65:263` generated, no old refs/interop, and final Video/Audio Prefix SHA-256 equality. Core raw maximum deltas were `2.384185791015625e-07` Video and `5.960464477539063e-08` Audio; final restored deltas are zero. No generated region is restored.
|
||||
- Both primary MP4s are finite 640x640 / 240 frames / 24 fps / 10.000 seconds. Nine-frame boundary contact sheets show no gross visual snap, collapse, persistent artifact, or scene discontinuity.
|
||||
- Added `tools/v36_r2_audio_boundary_analyze.py`. With Audio Seam OFF, the exact 124-frame cumulative boundary is the maximum decoded single-sample jump in both files: Reference `0.352289`, Masked `0.287521`. Masked has smaller DC/RMS/Peak discontinuities but the native hard cut is not described as click-free.
|
||||
- Added a cache-only Audio Seam Auto probe. It reassembled the accepted Masked latents in `7.23s` without Sampling; boundary jump became `0.005774` (about `98.0%` lower), global rank fell from maximum to `37.4` percentile, and DC jump became `0.001103`. Existing Audio Seam Auto therefore handles the observed cut; no new seam algorithm was added.
|
||||
- R2-1 status: GPU structure/interoperability **PASS**, raw Seam-OFF audio cut documented, existing Audio-Seam-Auto path PASS, Production still HOLD. R2-2 22f / 37T and broader acceptance remain pending. Evidence is under `D:\Codex\_test_results\ComfyUI-H3-Continuum\v36-r2-gate1-20260826`; commit/push were not performed.
|
||||
|
||||
## 2026-08-26 - V3.6-R2-0 Joint AV masked-prefix CPU/static gate
|
||||
|
||||
- Preserved all uncommitted V3.6-R1 work and kept V3.5.3 public/default behavior frozen. Added only the private `masked_av_prefix_39_v1` research transport; no public node/schema, `v2/sampling.py`, V3.4, seed, Plan, trim, Decode, Assembly, Seam, or Run Storage contract changed.
|
||||
- Generalized `v3/masked_continuation.py` so R1 still protects only 22f / 7T video, while R2-0 copies and masks both a 39f / 12T video tail and a duration-matched 65T audio tail. Sampling-finalization restore is limited to the two protected prefixes; generated video and audio regions remain sampler-owned.
|
||||
- Added a Strong-39-only sequence branch using a 158-frame physical target (`47T` video / `263T` audio), 39-frame trim, and 119-frame net output. It emits neither old Context Reference rows nor the old Spectrum interop hint and does not reapply the initial I2VA First Image after chunk 1.
|
||||
- Extended the private WAN diagnostic wrapper to capture audio-prefix hashes/equality, audio mask regions, source-tail matching, native audio-grid offset, and final physical-group AV prefix pairs. It remains diagnostic-only and is not a public V3.5.3 node contract.
|
||||
- Added CPU regressions for exact shapes, zero/one masks, source/target immutability, invalid audio shapes/NaN, protected-prefix restoration, generated-region preservation, First Image isolation, absence of old refs/interop, and the integrated two-chunk sequence path.
|
||||
- The 39f video window and 65 audio ticks are both 1.625 seconds, but native H3 source/target audio-grid phase is not silently assumed to be zero: 124f / 207T is `+1/3` tick and 158f / 263T is `-1/3` tick.
|
||||
- Validation: changed-file `py_compile` PASS; focused masked/temporal tests `21 passed`; full pytest `507 passed`; only the known `pynvml` FutureWarning remains. GPU R2-1 and runtime synchronization were deliberately not started in this CPU/static gate.
|
||||
- Accepted snapshots: `pre-v36-r2-0-joint-av-source-accepted-20260826_203022` and `pre-v36-r2-0-joint-av-wan-runtime-accepted-20260826_203022`. Commit/push were not performed; the existing WAN backend was not stopped.
|
||||
|
||||
## 2026-08-26 - V3.6-R1 masked video-prefix continuation PoC
|
||||
|
||||
- Kept V3.5.3 `reference_context_v1` as the public/default continuation transport and added only a private V3.6-R1 switch for testing. No public widget or Production node was added; V3.4, public V3.5.3 schemas, Sampling implementation, seeds, Plan, trim, Decode, Assembly, Seam, and Run Storage remain unchanged.
|
||||
- Added `v3/masked_continuation.py` for 22-frame / 7T video-prefix target construction and mask creation. The masked route copies the video tail into the next 42T target, uses video mask `0` for the seven protected slots and `1` afterward, keeps the 235T audio target empty with an all-one mask, and omits both old Context Reference conditioning and its Spectrum interop hint.
|
||||
- Found that Core v0.34.0 preserved the prefix inside denoising but the raw final float32 integration introduced a maximum difference of `2.38418579101562e-07`. Added a narrowly scoped post-Sampling restore of the protected prefix only. Finalized prefix equality is exact by `torch.equal`, max difference, and SHA-256; no generated video/audio region is restored.
|
||||
- Added dedicated CPU regressions, a test-runtime diagnostic node, and `tools/v36_r1_api_runner.py`. Focused validation is `8 passed`; full pytest is `500 passed`; `py_compile` and `git diff --check` pass. Synced only the changed Production/diagnostic files to ComfyUI_WAN and verified source/runtime SHA-256 equality.
|
||||
- After user-authorized queue clear/interrupt, completed three normal-API A/B pairs at 640x640 / 2x5-second T2VA / 20 steps / Balanced 22. Audio Continuity, Run Storage, Spectrum, Sol, Seam, First/Last, References, and Video Guide were off; Sage Attention was enabled by the server. Every masked run passed 7T prefix exactness, mask/shape/seed/finite checks, empty second-group refs, and absent old interop. All six outputs are 240 frames at 24 fps / 10.000 seconds.
|
||||
- Contact sheets at frames 112, 118, 122, 123, 124, 125, 128, 136, and 148 show no gross boundary snap, direction reversal, persistent artifact, or collapse for either route. The masked route did not show a consistent visual advantage over the accepted reference-context route across the three seeds. Decision: R1 structure PASS; visual improvement unproven; do not promote it as the Production default.
|
||||
- Evidence roots: `v36-r1-gate1-postrestore-20260826`, `v36-r1-gate2-seed238985343135587`, `v36-r1-gate2-seed238985343135588`, and `v36-r1-boundary-sheets-20260826` under `D:\Codex\_test_results\ComfyUI-H3-Continuum`. Output MP4s remain under `D:\output\video\comfy_video\video` with `V36_R1_*` names. Commit/push were not performed.
|
||||
- Derived a simplified, annotated A/B workflow from `MiniMax_H3_Continuum_V35 (3).json`: `D:\Codex\_test_workflows\ComfyUI-H3-Continuum\MiniMax_H3_Continuum_V36_R1_MaskedPrefix_AB.json`. Removed disconnected external AddGuide/Noise Mask/AV split-concat nodes and all non-R1 features; retained the exact Model/SIGMAS/diagnostic Sampler/Core Decode/Assembly/Save route. Static graph validation found 22 unique nodes, 24 reciprocal/type-correct links, three groups, zero errors, registered diagnostic type, and default `masked_video_prefix_v1`. Companion explanation is `V36_R1_WORKFLOW_GUIDE.md`; no GPU generation was rerun for workflow serialization.
|
||||
- Ran the formal R1 Stress Gate as 3x5-second Strong Motion I2VA at 640x640, Balanced 22, three fixed seeds, Prompt/CLIP cache HIT, Audio Continuity/Run Storage/Spectrum/Sol/Seam OFF, and `--disable-pinned-memory`. The order alternated `Reference -> Masked`, `Masked -> Reference`, `Reference -> Masked`.
|
||||
- Diagnostics caught an A/B fairness defect before acceptance: the private masked branch reapplied the original First Image on chunks 2/3 while the reference branch removed it. Interrupted/excluded those preliminary runs, restricted First Image to chunk 1 in `masked_video_prefix_v1`, and added CPU coverage. Public V3.5.3 behavior was not changed.
|
||||
- The accepted normal API pairs were `719.800/675.284s`, `693.449/641.506s`, and `711.800/672.341s` for Reference/Masked. Masked API time was lower by `6.185%`, `7.491%`, and `5.544%`; continuation Sampling was lower by `9.471%`, `12.029%`, and `8.309%`, with paired medians `6.185%` and `9.471%` respectively.
|
||||
- Added real Core `PackedLayout` capture to the diagnostic node. Reference continuation uses `20,569` rows including `2,800 ref_img` rows; Masked uses `17,769` rows and no continuation `cond`/`ref_img` rows. A CUDA-synchronized pair confirmed continuation Sampling `464.97 -> 437.10s` (`5.995%` faster), so the signal is not a host-asynchrony artifact.
|
||||
- All six accepted videos are 640x640, 360 frames, 24 fps, and 15.000 seconds. Every masked finalized prefix is bit-exact at both boundaries. Contact sheets show no gross snap, reversal, persistent artifact, or collapse. At the second boundary Masked had lower optical-flow jerk and lower absolute luma/saturation drift for all three seeds; flow-magnitude continuity remained mixed, so the quality result is promising but not universal proof.
|
||||
- Stress evidence is under `D:\Codex\_test_results\ComfyUI-H3-Continuum\v36-r1-stress-gate-20260826\normal_fair` and `sync_fair`. The sibling `normal` runs are pre-fairness diagnostics and explicitly invalid for acceptance. Added reusable `tools/v36_r1_stress_gate_runner.py` and `tools/v36_r1_stress_gate_analyze.py`.
|
||||
- Final validation: focused masked-prefix regression `5 passed`; canonical full pytest `500 passed`; changed-file `py_compile`, V3.5.3 runtime verifier, `git diff --check`, source/WAN hashes, and empty queue PASS. R1 Stress Gate is PASS; Production remains HOLD pending separately scoped R2/Reference/Terminal/Run Storage work. The WAN backend was intentionally left running per user instruction. Commit/push were not performed.
|
||||
|
||||
## 2026-08-26 - V3.5.3 distribution-integrity maintenance hotfix
|
||||
|
||||
- Kept published V3.5.2 generation and optimization behavior frozen; no Sampling, Conditioning payload, Terminal Merge, Assembly, Seam, Run Storage, Prompt/CLIP cache, Video Guide, V3.4, node, or socket contract was changed.
|
||||
|
||||
Reference in New Issue
Block a user