Release v3.6.0 masked AV continuation
This commit is contained in:
+100
-3
@@ -10,7 +10,104 @@
|
||||
- E2-8A Contract, E2-8B1 CPU/static validation, E2-8B2 external LBH GPU PoC, and E2-8C Second Pass GPU acceptance are formally PASS. The E2-8C CPU baseline is `260 passed`.
|
||||
- External LATENT nodes may discard custom dictionary metadata. Identically shaped group reordering therefore cannot be cryptographically detected; count, position, T, geometry, and audio-shape checks are enforced.
|
||||
|
||||
Updated: 2026-08-26
|
||||
Updated: 2026-08-27
|
||||
|
||||
## V3.6.0 public release implementation (2026-08-27)
|
||||
|
||||
- Added the new public `H3ContinuumSamplerV36` / `H3 Continuum Sampler V3.6` without replacing or redirecting V3.5.3/V3.4 nodes. Its final required widget is `Continuation Backend` with only `Standard` and `Compatibility`; internal transport identifiers are not exposed.
|
||||
- `Standard` selects accepted Masked AV when Audio Continuity is ON and accepted Masked Video when it is OFF. `Compatibility` selects the original Reference Context route. Existing V3.5.3 remains on Reference Context with its prior schema and widget sequence.
|
||||
- Added `examples/workflows/MiniMax_H3_Continuum_V36.json` as the recommended template, retaining the V3.5 template unchanged. Version/package metadata and English/Japanese release documentation are updated to 3.6.0; V3.5 Second Pass, Hi-Res Fix, Conditioning Bridge, V3.5 Assembly, and older workflows remain available.
|
||||
- The public-node GPU Smoke used `H3ContinuumSamplerV36` directly through the normal API, not the private diagnostic node: Standard backend, 2x5-second I2VA + Reference Image + Reference Audio, 384x384, four steps, Audio Continuity ON, Audio Seam Auto. Prompt `1cff2ce7-6587-4d71-903d-b04bb304fdd5` completed successfully in `65.21s`; the report confirms a 22f/37T masked AV target prefix on chunk 2.
|
||||
- Smoke output is `D:\output\video\comfy_video\video\V36_Release_Smoke_Reference_AV_Standard_00001_.mp4`: finite 384x384, 240 frames, 24 fps, 10.000-second video, and 32 kHz stereo 10.000-second audio. Evidence is `D:\Codex\_test_results\ComfyUI-H3-Continuum\v360-release-smoke\20260827_172835`.
|
||||
- Source focused release validation is `81 passed`; final full source validation is `527 passed` with only the known `pynvml` FutureWarning. V3.6.0 WAN runtime verifier, JavaScript syntax, changed-file `py_compile`, the 178-entry Manifest, README link validation, and source/runtime SHA-256 equality for 12 Production files all pass. The live API queue is empty and the V3.6 backend remains running.
|
||||
- Accepted snapshots are `pre-v360-public-release-accepted-20260827_171442` (212 Git-visible source files) and `pre-v360-public-runtime-wan-20260827_172531` (the five runtime files changed for public registration/version/UI). Final repository validation is complete; commit/push/release are the remaining publication actions.
|
||||
|
||||
## V3.6 Production Integration Gate (2026-08-27)
|
||||
|
||||
- PIG-0 through PIG-5 are complete. The Masked AV backend and V3.6 public/default promotion are **PASS**. During the gate, published/default V3.5.3 remained `reference_context_v1` and no public schema changed; the separately authorized V3.6 UI/version implementation is recorded above. V3.5.3 and V3.4 behavior remains unchanged.
|
||||
- Run Storage now scopes only nondefault continuation transports. `reference_context_v1` retains its exact legacy identity/revision behavior, while `masked_av_prefix_22_v1` records `continuation_transport` in both the identity and Sampling Contract. Reference and Masked chunks therefore cannot be mixed.
|
||||
- The integration gate also found and fixed a pre-existing resume defect: temporary Run Storage Sessions did not restore Terminal Merge `execution_semantics`, so an apparently valid prefix was discarded before Sampling. Resume Sessions now carry the stored execution semantics, and the reported reuse count is synchronized after atomic terminal-pair pruning.
|
||||
- Real-cache PIG-2 PASS evidence is under `D:\Codex\_test_results\ComfyUI-H3-Continuum\v36-production-integration\v36_pig2_masked_20260827_164303`. Initial Save used `0 reused / 3 generated`; Auto Resume used `3 / 0` with zero Sampling calls; Regenerate From Chunk 2 and Chunk 3 each used `1 / 2` and exactly one `77T Video / 433T Audio` terminal physical Sampling call. Neither request reused or sampled only one logical half of the terminal pair.
|
||||
- PIG-1 CPU coverage proves that Reference Image and standalone Reference Audio remain present exactly once in every normal and Terminal physical group under Masked AV, with no old continuation Reference block. PIG-3 then completed a representative 2x5-second I2VA + Reference Image + Reference Audio GPU run at 384x384 / 4 steps. Both physical groups contained image/audio references, protected Video/Audio prefixes finalized bit-exact, and the final output is finite 240 frames / 24 fps / 10.000 seconds with 10.000-second stereo audio.
|
||||
- PIG-3 evidence is `D:\Codex\_test_results\ComfyUI-H3-Continuum\v36-production-integration\pig3-reference-masked`; output video is `D:\output\video\comfy_video\video\V36_PIG3_Reference_Image_Audio_Masked_AV_00001_.mp4`. At the retained 124-frame boundary with existing Audio Seam Auto, the sample jump is `0.004261`, below the local median (`0.005129`) and only `0.0876x` the local p99; PCM is finite. This representative numeric gate is PASS, but prior R2 seeds show mixed Seam-Auto improvement, so no universal click-removal claim is made without listening/generalization.
|
||||
- PIG-5 policy is fixed: V3.6 uses Masked AV as the standard continuation backend and retains Reference Context only as an Advanced compatibility fallback. Raw transport implementation names are not exposed in the UI. Run Storage identity remains transport-scoped, Terminal Merge regeneration remains atomic per physical group, and Audio Seam Auto remains an independently evaluated post-Sampling feature. This gate preceded the separately recorded V3.6 public UI/default implementation above.
|
||||
- The accepted PIG-3 output passed subjective listening: no click, dropout, unnatural stereo positioning, or other audible boundary defect was reported. PIG-5 and V3.6 public-promotion eligibility are therefore **PASS** without reopening the completed structural, cache, Reference, Terminal Merge, or Audio Seam numeric gates.
|
||||
- Final validation is `61 passed` focused and `521 passed` full pytest with only the known `pynvml` FutureWarning. Changed-file `py_compile`, V3.5.3 WAN runtime verification, source/runtime SHA-256 equality for `run_storage.py`, `v2/sequence.py`, and the diagnostic node, `git diff --check`, and empty API queue all pass.
|
||||
- Accepted pre-change snapshots are `pre-v36-production-integration-pig0-accepted-20260827_161517` and `pre-v36-production-integration-pig2-diagnostic-accepted-20260827_162943`. Commit/push were not performed, and the ComfyUI_WAN backend remains running.
|
||||
|
||||
## V3.6-R2 FL2VA Long Terminal Merge multi-seed GPU gate (2026-08-27)
|
||||
|
||||
- The shortened final gate is **PASS for the private GPU research path**. Published/default V3.5.3 remains `reference_context_v1`; no public node/schema, saved-workflow contract, V3.4 behavior, Sampling implementation, Assembly, Seam, or Run Storage behavior changed.
|
||||
- Accepted evidence consists of Seed `238985343135606` full Gate A, Seed `238985343135607` full `reference_context_v1` versus `masked_av_prefix_22_v1` comparison plus cache-only Audio Seam OFF derivatives, and Seed `238985343135608` Masked AV structure/practical acceptance. The redundant Seed 608 Reference run was intentionally omitted.
|
||||
- Every measured run retained 3 logical chunks and 2 physical groups. Group 1 is `124f / Video 37T / Audio 207T`; Terminal Group 2 is logical `[2,3]`, `paired_timeline_v1`, `terminal_merged=true`, trim `22f`, `260f / Video 77T / Audio 433T`. Physical seeds were `15439951268068580545`, `4181078118376080811`, and `6961917749643595863` for Seeds 606–608. Final outputs are finite `640x640`, `360f`, `24fps`, `15.000s`, with 15.000-second audio.
|
||||
- Seed 606 and 607 Group 1 Video/Audio hashes are identical between Reference and Masked routes. For all three Masked runs, Core raw protected-prefix drift was at most `2.384185791015625e-07` Video and `5.960464477539063e-08` Audio; scoped final restore produced bit-exact 7T Video and 37T Audio prefixes with SHA-256 equality and maximum difference `0`. Generated regions were not restored.
|
||||
- Terminal Group 2 PackedLayout was `35,865` rows for Reference and `32,991` for Masked, removing exactly `2,874` Reference rows (`8.01%`). CUDA-synchronized Group 2 Sampling was `574.134 -> 529.933s` for Seed 606 and `560.782 -> 544.413s` for Seed 607. Both paired comparisons favor Masked; paired percentage median is `5.31%`. Reference/Masked time medians are `567.458s` and `544.413s`, respectively (`4.06%` lower by median comparison). Seed 608 Masked was `549.704s`.
|
||||
- Across the five main runs, monitored process peak RSS median was `9.674 GiB`. Maximum observed whole-device usage was `15.646 GiB`; this is an order-sensitive device-used monitor, not only Continuum allocation. Terminal-group CUDA allocator peak was about `4.06 GiB` for Reference versus `3.80 GiB` for Masked. No OOM, NaN/Inf, or retained-GPU-prefix regression occurred.
|
||||
- Contact sheets covering nominal 5/10/15-second positions and actual retained boundaries `124f/243f` show no obvious flash, pose collapse, direction reversal, or temporal break in Seeds 606–608. Decoded audio is finite, exactly 15 seconds, and has no >=200 ms dropout below -50 dB. Existing Audio Seam Auto strongly improved Seed 606 Masked jumps (`0.035014 -> 0.002018` at 124f; `0.004495 -> 0.003130` at 243f), but Seed 607 was mixed (`0.059640 -> 0.133805` at 124f; `0.021297 -> 0.019704` at 243f). Therefore no universal click-removal claim is made from numerical jump alone.
|
||||
- Final validation is `516 passed` with only the known `pynvml` FutureWarning, after rerunning with a fresh writable basetemp because the inherited global pytest temp root denied access. V3.5.3 WAN runtime verification, source/runtime SHA-256 for seven synchronized Production/diagnostic files, `git diff --check`, and empty API queue pass. The backend remains running as requested.
|
||||
- **Production promotion: HOLD.** FL2VA/Terminal Merge structure, multi-seed execution, integrity, and performance are accepted, but Reference conditioning variants, Run Storage/resume identity, public UI/default-policy decisions, and broader audible Audio Seam acceptance remain outside this gate. Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v36-fl2va-combined-20260826\gate-a` and `...\r2-short`. Pre-change snapshot: `pre-v36-r2-short-gate-20260827_000840`. Commit/push were not performed.
|
||||
|
||||
## V3.6 shared chunk-duration range (2026-08-27)
|
||||
|
||||
- V3 Production, V3.4/V3.5 inherited schemas, the V2 compatibility sampler/facade, and Prompt Plan Preview now share `chunk_seconds`: default `5.0`, minimum `4.0`, maximum `30.0`, step `0.1`.
|
||||
- One `chunk_seconds` value still applies to every logical chunk. Per-chunk variable durations were not added; Assembly, Timeline, Continuation, and Run Storage retain the shared-duration contract and existing workflow inputs remain compatible.
|
||||
- `5–15s` remains the recommended and validated range. `15–30s` is usable rather than hidden/experimental, with a shared tooltip warning that high resolution can substantially increase VRAM use and processing time.
|
||||
- Prompt Plan creation and validation accept `30.0` and reject values above it. Timeline mapping was verified for `[0-30s]` / `[30-60s]`. Temporal extension tests now cover the complete `4.0–30.0s` UI range and preserve the existing native H3 `17k+5` grid.
|
||||
- The only remaining `max: 15.0` is the separate legacy V1 `H3ContinuumJoin.extend_seconds` control; it is not `chunk_seconds`, is not used by the V2/V3 Production path, and was intentionally left unchanged.
|
||||
- CPU/static acceptance is PASS: focused duration and inherited-schema validation `89 passed`; canonical full pytest `516 passed`; changed-file `py_compile`, `git diff --check`, runtime verifier, source/WAN SHA-256, live API schema, and empty queue all pass. ComfyUI_WAN was restarted with Sage Attention and `--disable-pinned-memory` and remains running.
|
||||
- A validation-only GPU launch used one T2VA chunk, `chunk_seconds=30.0`, `384x384`, one Sampling step, and fixed seed `238985343135609`. API prompt `9e665f85-b32f-4b17-9734-219fb284e087` was accepted with no node errors and entered H3 model initialization/Sampling without duration, shape, Prompt Plan, immediate OOM, or NaN/Inf errors. Per user scope, it was interrupted after `31.87s` rather than decoded or saved; this is an execution-entry smoke PASS, not a completed 30-second quality/performance acceptance. Queue returned to `0/0`.
|
||||
- The previously prepared FL2VA joint multi-seed run was cancelled at the user's request. Only its one-step cache warm-up completed; the first measured run was interrupted before node execution and is not acceptance evidence.
|
||||
- Accepted snapshots are `pre-v36-chunk-duration-30-source-20260826_234716` (209 Git-visible source files, zero missing/mismatch) and `pre-v36-chunk-duration-30-wan-runtime-20260826_235423` (four exact runtime files). Commit/push were not performed.
|
||||
|
||||
## V3.6-R2-2 Balanced 22f / 37T Joint AV GPU gate (2026-08-26)
|
||||
|
||||
- R2-2 is **PASS as a one-seed GPU structure/performance gate with the existing Audio Seam Auto path**, not a Production promotion. The accepted pair used T2VA 2x5 seconds, 640x640, Balanced 22, Audio Continuity ON, fixed seed `238985343135605`, Prompt/CLIP cache HIT, CUDA synchronization, Sage Attention, `--disable-pinned-memory`, Exact Duration ON, and no First/Last/Reference/Video Guide/Run Storage/Spectrum/Sol/Video Seam.
|
||||
- Added only the private `masked_av_prefix_22_v1` research transport. It preserves the existing Balanced contract: Video prefix `22f = 7T = 0.916667s`, Audio prefix `37T = 0.925s`, an intentional mask-boundary asymmetry of `1/120s = 8.333ms`, target `141f / Video 42T / Audio 235T`, trim `22f`, and net new `119f`. Public V3.5.3/V3.4 schemas and defaults remain unchanged.
|
||||
- Masked finalization is exact: source Video tail and finalized target Video prefix are bit-exact with max diff `0`; the same holds for the Audio 37T tail/prefix. Core raw output before the scoped restore differed by at most `2.384185791015625e-07` for Video and `5.960464477539063e-08` for Audio. Only the protected prefix regions are restored; generated regions remain sampler-owned.
|
||||
- Reference continuation used `20,205` packed rows (`2,800` video-reference rows plus `74` audio-reference rows); Masked AV used `17,331` rows with no reference block or old Spectrum interop hint. The `2,874`-row reduction is `14.22%`. Synchronized continuation Sampling was `226.373s -> 206.437s`, an `8.81%` reduction; whole API time was `435.907s -> 420.752s`, a `3.48%` reduction. This is one-seed evidence, not a universal performance claim.
|
||||
- Both routes produced finite `640x640`, `240f`, `24fps`, `10.000s` outputs. Eight-frame boundary contact sheets show no obvious pose snap, direction reversal, camera collapse, NaN/Inf, or masked-prefix visual regression.
|
||||
- With Audio Seam OFF, Reference boundary jump was `0.082725`; Masked AV was worse at `0.268163` (`7.57x` its local p99 and the `99.9989` file-wide percentile). Reassembling the same cached Masked latents with existing Audio Seam Auto took `7.10s` without Sampling and reduced the jump to `0.029944` (about `88.83%` lower, `0.86x` local p99, `94.31` percentile). RMS ratio also moved `1.12196 -> 0.97724`. R2 adds no new Audio Seam algorithm; raw Seam-OFF 22f/37T is not accepted as click-free.
|
||||
- Evidence root: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v36-r2-gate2-20260826`; accepted MP4s and the cached Audio-Seam-Auto derivative are under `D:\output\video\comfy_video\video`. Focused tests are `24 passed`; final full pytest is `510 passed`; changed-file `py_compile`, `git diff --check`, source/runtime SHA-256, runtime verifier, and empty queue all pass. Production remains **HOLD** pending multi-seed/audio listening generalization, Reference/FL2VA/Terminal Merge/Run Storage integration, public UI decisions, and final Production acceptance.
|
||||
|
||||
## V3.6-R2-0 Joint AV masked-prefix CPU/static gate (2026-08-26)
|
||||
|
||||
- R2-0 is implemented as a private research transport, `masked_av_prefix_39_v1`; the public/default V3.5.3 transport remains `reference_context_v1`. No public widget, Node ID, saved-workflow schema, `v2/sampling.py`, seed, Plan, trim, Decode, Assembly, Seam, Run Storage, or V3.4 contract changed.
|
||||
- The exact-duration research case uses Strong 39-frame Continuity. A 39-frame video tail maps to 12 video latent slots, and the same 1.625-second duration maps to 65 audio latent ticks. The next physical target is 158 frames / video `47T` / audio `263T`, trims 39 frames, and retains 119 new frames.
|
||||
- The target copies the prior finalized video tail into video `[0:12]` and audio tail into audio `[0:65]`. Full masks preserve those prefixes with zero and generate only video `[12:47]` / audio `[65:263]` with one. Old Context Reference conditioning and the Spectrum reference-context hint are not emitted. The original First Image remains attached only to chunk 1.
|
||||
- After Sampling, Continuum restores only the protected video and audio prefixes from the normalized pre-sampling target. CPU regressions require bit-exact video/audio prefix equality and prove that generated video/audio regions are not restored or overwritten.
|
||||
- The 39f and 65T durations are equal, but the native H3 audio grid can retain a signed source phase offset. The validated 124-frame source is `+1/3` audio tick (about `+8.33 ms`) and a 158-frame target is `-1/3` tick; this is recorded explicitly rather than described as zero-phase alignment.
|
||||
- CPU/static R2-0 is **PASS**: focused masked/temporal validation `21 passed`, canonical full pytest `507 passed`, and changed-file `py_compile` pass with only the known `pynvml` FutureWarning. GPU R2-1 (`Reference Context 39f` versus `Masked AV Prefix 39f`), audio-boundary measurements, runtime synchronization, and Production promotion remain pending.
|
||||
- Accepted pre-change snapshots are `pre-v36-r2-0-joint-av-source-accepted-20260826_203022` (571 files, SHA-256 mismatch 0, source revision `67a0d80117bd8332d2768d093dff343001c87d60`) and `pre-v36-r2-0-joint-av-wan-runtime-accepted-20260826_203022` (three exact runtime files). Commit/push were not performed.
|
||||
|
||||
## V3.6-R2-1 39f / 65T Joint AV GPU structure gate (2026-08-26)
|
||||
|
||||
- R2-1 is **PASS as a GPU structure/interoperability gate**. It is not a Production promotion. The accepted pair used T2VA 2x5 seconds, 640x640, Strong 39, Audio Continuity ON, the same fixed prompt/seed `238985343135604`, Prompt/CLIP cache HIT, CUDA synchronization around each Sampling call, `--disable-pinned-memory`, Sage Attention, and no First/Last/Reference/Video Guide/Run Storage/Spectrum/Sol/Video Seam. Reference ran before Masked after one excluded warm-up.
|
||||
- `reference_context_v1` emitted Core's combined `video_audio` Reference block. Its continuation PackedLayout was `24,317` rows: target text `61`, audio `526`, video `18,800`, plus `4,800 ref_img` and `130 ref_audio` rows. `masked_av_prefix_39_v1` emitted no old refs/interop and used `19,387` rows: the identical target rows only. The measured reduction is exactly `4,930` rows / `20.27%` relative to Reference.
|
||||
- CUDA-synchronized Chunk 2 host/Sampling time was `332.460s` for Reference and `241.888s` for Masked (`27.24%` lower). API wall time was `556.058s` versus `465.925s` (`16.21%` lower). This one-seed Gate is an explanatory performance signal, not a generalized speed claim.
|
||||
- Masked target construction and masks passed on GPU: Video `[0:12]=0`, generated Video `[12:47]=1`, Audio `[0:65]=0`, generated Audio `[65:263]=1`; source tails matched both target prefixes. Core raw integration differed by at most `2.384185791015625e-07` for Video and `5.960464477539063e-08` for Audio. The finalized physical-group prefixes are bit-exact by SHA-256 and maximum difference `0` for both streams; generated regions are not restored.
|
||||
- Both outputs are finite `640x640`, 240 frames, 24 fps, and 10.000 seconds. Boundary contact sheets at frames `112,118,123,124,125,130,136,148,160` show no gross snap, collapse, persistent artifact, or identity/background discontinuity in either route.
|
||||
- With Audio Seam OFF, the exact decoded boundary sample is the global maximum single-sample jump for both routes: Reference `0.352289`, Masked `0.287521`. Masked is better but the raw hard boundary is not accepted as intrinsically click-free. The same cached Masked AV latents reassembled with the existing Audio Seam Auto in `7.23s` without re-running Sampling; boundary jump fell to `0.005774` (about `98.0%` lower and only the `37.4` percentile of file-wide sample differences), while DC jump fell `0.002789 -> 0.001103`. Therefore the existing Production seam path resolves the observed hard cut; R2 adds no new Audio Seam algorithm.
|
||||
- Chunk 2 allocator peak was lower for Masked (`2,407,722,872` allocated / `2,785,017,856` reserved bytes) than Reference (`2,950,494,614` / `3,590,324,224`). Whole-process/device monitor peaks are order-sensitive and are not accepted as a memory claim from this one pair.
|
||||
- Evidence root: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v36-r2-gate1-20260826`; accepted MP4s are `V36_R2_Gate1_reference_39f_seed238985343135604_00001_.mp4`, `V36_R2_Gate1_masked_av_39f_65t_seed238985343135604_00001_.mp4`, and the cached Audio-Seam-Auto derivative under `D:\output\video\comfy_video\video`. The backend remains running. R2-2 22f / 37T, multi-seed/audio-quality generalization, References, Terminal Merge, Run Storage, public UI, and Production promotion remain pending.
|
||||
|
||||
## V3.6-R1 masked video-prefix continuation PoC (2026-08-26)
|
||||
|
||||
- V3.6-R1 is a private experimental continuation transport only. The published V3.5.3 path remains `reference_context_v1`; no public widget, Production Node ID, saved-workflow schema, V3.4 contract, Sampling implementation, seed derivation, Plan, trim, Decode, Assembly, Seam, or Run Storage contract was changed.
|
||||
- `masked_video_prefix_v1` copies the previous chunk's normalized 22-frame / 7T video tail into the next 42T video target, applies a full video mask with prefix `0` and generated region `1`, leaves the 235T audio target empty with an all-`1` mask, emits neither the old `minimax_refs` video context nor the Spectrum reference-context hint, and retains the existing 141-frame physical / 22-frame trim / 119-frame net contract.
|
||||
- Core v0.34.0 honors the zero-mask prefix during model calls, but its raw final integrated float32 sampler output differed from the input prefix by up to `2.38418579101562e-07`. Continuum therefore restores only the protected target prefix after Sampling. The finalized physical latent is bit-exact (`torch.equal`, maximum absolute difference `0`, and SHA-256 match); generated video and audio regions are not overwritten.
|
||||
- CPU/static acceptance is PASS: focused validation `8 passed`; canonical full pytest `500 passed` with only the known `pynvml` FutureWarning; changed-file `py_compile` and `git diff --check` pass.
|
||||
- GPU Gate 1 and Gate 2 ran three fixed-seed A/B pairs through the normal ComfyUI_WAN API at `127.0.0.1:8188`: 2x5-second T2VA, 640x640, 20 steps, Balanced 22, Audio Continuity OFF, Run Storage OFF, no First/Last/Reference/Video Guide, Spectrum/Sol OFF, Seam OFF, and Sage Attention enabled globally. Every run preserved two physical samples (`124f` then `141f`, trim `22`, net `119`), matching derived seeds and chunk-1 video/audio hashes. All six outputs are finite 640x640 / 240-frame / 24 fps / 10.000-second MP4s.
|
||||
- All masked runs report prefix exactness, zero/one video mask regions, all-one audio mask, empty second-group `minimax_refs`, and no old interop hint. API elapsed A/B pairs were `432.67/412.73`, `409.98/386.09`, and `408.62/386.56` seconds; these single ordered pairs are diagnostic observations, not an accepted speed claim.
|
||||
- The initial 2x5 T2VA contact-sheet review found no clear pose snap, direction reversal, persistent artifact, or temporal collapse in either transport, but also no consistent visual superiority. The later 3x5 Strong Motion I2VA Stress Gate adds a positive signal without yet justifying a public/default replacement.
|
||||
- Before the formal I2VA gate, diagnostics found that the private masked branch reapplied the original I2VA First Image on continuation chunks while `reference_context_v1` correctly removed it under `POLICY_REPLACE`. Those preliminary runs are invalid and excluded. The private branch now attaches the First Image only to chunk 1, so the accepted A/B differs only in continuation transport.
|
||||
- Formal Stress Gate conditions were 3x5-second I2VA, 640x640, Balanced 22, three fixed seeds, Audio Continuity/Run Storage/Spectrum/Sol/Seam OFF, Prompt/CLIP cache HIT, `--disable-pinned-memory`, and alternating order `R->M`, `M->R`, `R->M`. All six normal runs are finite 640x640 / 360-frame / 24 fps / 15.000-second MP4s.
|
||||
- Masked continuation was faster for all three seeds. API reductions were `6.185%`, `7.491%`, and `5.544%` (paired median `6.185%`). Continuation Sampling reductions were `9.471%`, `12.029%`, and `8.309%` (paired median `9.471%`; reference `466.424s`, masked `426.798s`). A separate CUDA-synchronized seed pair still reduced continuation Sampling from `464.97s` to `437.10s` (`5.995%`), proving that the signal is not only asynchronous host timing.
|
||||
- Actual Core packed layouts explain the result: continuation `reference_context_v1` is `20,569` rows (`2,800` `ref_img` rows), while `masked_video_prefix_v1` is `17,769` rows with no `cond` or `ref_img` rows. This is `2,800` fewer rows / `13.61%` below the reference layout. Peak process RSS and observed device memory remained comparable; no memory regression was accepted or claimed.
|
||||
- Every masked boundary retained the finalized 7T prefix bit-exact by SHA-256 and maximum difference `0`. Boundary contact sheets found no gross snap, direction reversal, persistent artifact, or collapse. At the second boundary, automated optical-flow jerk and absolute luma/saturation drift were lower for Masked in all three seeds; flow-magnitude continuity was mixed, so this is an improvement signal rather than proof of universal visual superiority.
|
||||
- R1 Stress Gate is therefore **PASS**, while Production promotion remains **HOLD**. It opened the separately specified R2 Audio-prefix research gate; the later R2-0 CPU/static result is recorded above. Feathering, References, Terminal Merge, public UI, Run Storage integration, and GPU R2 acceptance remain unimplemented V3.6 research.
|
||||
- Stress evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v36-r1-stress-gate-20260826\normal_fair` and `sync_fair`; `analysis.json` and `contact_sheets` contain the aggregate metrics and boundary frames. Invalid pre-fairness runs under the sibling `normal` directory are retained only as diagnostic history and must not be cited. Accepted Stress snapshots are `pre-v36-r1-stress-gate-source-accepted-20260826_174839`, `pre-v36-r1-stress-gate-wan-diagnostics-accepted-20260826_174839`, and `pre-v36-r1-stress-i2va-fairness-wan-runtime-20260826_182905`.
|
||||
- Final validation after the Stress Gate is `500 passed`; changed-file `py_compile`, V3.5.3 runtime verification, `git diff --check`, source/runtime SHA-256 for four R1 Production files plus the diagnostic node, and empty API queue all pass. The ComfyUI_WAN backend intentionally remains running after acceptance.
|
||||
- Reusable private A/B workflow: `D:\Codex\_test_workflows\ComfyUI-H3-Continuum\MiniMax_H3_Continuum_V36_R1_MaskedPrefix_AB.json` (22 nodes, 24 links, three annotated groups, SHA-256 `0964D55DDE1671E0BB0CA04FD3DD1F0AF82A4173FB9E6D9CE75900FC9B6981CE`). It is a simplified derivative of the user-supplied V3.5 workflow and requires the ComfyUI_WAN-only diagnostic node; it is not a public template.
|
||||
|
||||
## V3.5.3 maintenance hotfix (2026-08-26)
|
||||
|
||||
@@ -21,7 +118,7 @@ Updated: 2026-08-26
|
||||
- Added regressions for all eight public workflow graphs and the Hybrid warning numbers. Focused validation is `30 passed`; canonical full pytest is `495 passed`, and the source/runtime V3.5.3 verifier passes native PackedLayout, Fixed 3x5 planning, and complete node registration.
|
||||
- The release Manifest now contains 173 files and verifies every recorded SHA-256, including the new public-workflow integrity regression.
|
||||
- Production files `reference.py`, `version.py`, and `metadata.ini` are synchronized to ComfyUI_W with exact SHA-256 equality. ComfyUI_WAN is excluded because that secondary runtime is an incomplete older copy and must not be partially relabeled as V3.5.3.
|
||||
- Accepted snapshots are `pre-v353-maintenance-hotfix-source-accepted-20260826_030732` (199 tracked files, revision `70cbbcd`) and `pre-v353-maintenance-hotfix-comfyui-w-accepted-20260826_030731` (14 present target files). Commit and push have not been performed.
|
||||
- Accepted snapshots are `pre-v353-maintenance-hotfix-source-accepted-20260826_030732` (199 tracked files, revision `70cbbcd`) and `pre-v353-maintenance-hotfix-comfyui-w-accepted-20260826_030731` (14 present target files). The maintenance release was published as commit `67a0d80117bd8332d2768d093dff343001c87d60`.
|
||||
|
||||
## V3.5.2 final release gate (2026-08-26)
|
||||
|
||||
@@ -142,7 +239,7 @@ Updated: 2026-08-26
|
||||
|
||||
## Baseline
|
||||
|
||||
- Release-candidate baseline: V3.5.3 maintenance hotfix (published baseline remains V3.5.2 until commit/push)
|
||||
- Published baseline: V3.5.3 maintenance hotfix (`67a0d80117bd8332d2768d093dff343001c87d60`)
|
||||
- Authoritative working repository: `D:\Codex\_git_push_work\ComfyUI-H3-Continuum`
|
||||
- Primary tested runtime: `D:\StabilityMatrix\Data\Packages\ComfyUI_W\custom_nodes\ComfyUI-H3-Continuum`
|
||||
- Secondary runtime: `D:\StabilityMatrix\Data\Packages\ComfyUI_WAN\custom_nodes\ComfyUI-H3-Continuum`
|
||||
|
||||
Reference in New Issue
Block a user