54 lines
4.2 KiB
Plaintext
54 lines
4.2 KiB
Plaintext
ComfyUI-H3-Continuum V3.7.0 package validation
|
|
|
|
Status: V3.7.0 release candidate. Conditioning Adapter and the internal RefineSchedule v1 are correctness-accepted without changing V3.6 Production defaults or the existing SIGMAS socket. Still Image Guide positioning, source ownership, Terminal Merge, Run Storage identity, and target-resolution reconditioning are correctness-accepted, but hard-anchor quality remains Experimental / Production HOLD. All V3.6/V3.5.3/V3.4 public node IDs and saved-workflow contracts remain available.
|
|
|
|
Required automated checks:
|
|
- python -m compileall -q .
|
|
- python -m pytest -q -p no:cacheprovider
|
|
- python tools/verify_runtime.py --comfy-root <runtime-path>
|
|
|
|
Validated runtime paths include:
|
|
- H3 Continuum Sampler V3.7 required-input parity with V3.6 and one optional Still Image Guide input
|
|
- Conditioning Adapter reconstruction of First/Last, continuation context, and owned Still Guide sources at target geometry
|
|
- RefineSchedule v1 External / Full / Tail / Partial contracts with exact source schedule slicing
|
|
- Internal Tail 6 at 704x704: six evaluations, decoded Video SSIM 0.982936 versus the external Tail 6 baseline, and bit-exact decoded Audio PCM
|
|
- 1152x1152 Still Guide source reconditioning correctness with first-pass Audio passthrough preserved
|
|
- Still Guide single-IMAGE validation and explicit source-role priority at First/Last endpoint collisions
|
|
- V3.7 compact frontend parity for project identity, hidden widgets, Run Storage controls, queue normalization, diagnostics, and permanent Reference Audio sockets
|
|
- V3.6.1 mixed Timeline/List Prompt diagnostics and fail-open normalization
|
|
- V3.6.1 non-Balanced Audio Continuity fallback routing before Run Storage identity creation
|
|
- V3.6.1 public-node Strong 39 / Audio Continuity fallback GPU Smoke on ComfyUI_W: 2x5 seconds, 576x576, 240 frames, 10.000-second video and audio
|
|
- H3 Continuum Sampler V3.6 Standard / Compatibility backend mapping without exposing internal transport identifiers
|
|
- Masked AV protected Video/Audio prefix bit-exact finalization
|
|
- Run Storage Save, Auto Resume, and Regenerate From with backend-scoped revisions and atomic Terminal Merge reuse
|
|
- Reference Image and Reference Audio retention in every Masked physical group
|
|
- 3x5 FL2VA Long Terminal Merge Masked GPU acceptance and PIG-3 subjective Audio PASS
|
|
- Shared 4.0-30.0 second chunk duration range; 5-15 seconds remains the recommended and validated range
|
|
- V3.5 refine-context contract and Advanced Second Pass Bridge
|
|
- Conditioning Bridge MODEL / CONDITIONING alignment per physical group without CONDITIONING-entry flattening
|
|
- Permanent Python-defined Reference Audio sockets, workflow save/reload widget alignment, and conditioning retention
|
|
- 1x5 T2VA, 3x5 T2VA, and 3x5 FL2VA Long Terminal Merge Second Pass GPU acceptance
|
|
- First-pass audio LATENT object passthrough through Second Pass
|
|
- H3 Continuum Assemble + Seam V3.5 RAM / Disk-backed output equivalence
|
|
- Auto backend selection with RAM and 9.49 GiB disk-backed real-runtime cases
|
|
- Exact Duration and Video Seam direct-write behavior
|
|
- Experimental 1x5 FL2VA Pixel/VAE Hi-Res Fix GPU acceptance at 576 to 1152
|
|
- Hybrid FL2VA + Reference 1x5 GPU acceptance through integrated 2x Hi-Res Fix and direct Advanced Second Pass
|
|
- Production sampler and Core VAE Decode assembly
|
|
- Reference Image 1-3
|
|
- Driving Audio absolute-time guide conditioning with original-source final output
|
|
- Persistent Video Guide Frames conditioning
|
|
- Video Guide Size: Efficient 0.4 MP / Balanced 0.6 MP / Match Output
|
|
- Run Storage save, checksum verification, and complete reuse
|
|
- Run Storage identity coverage for Driving Audio and Video Guide Frames
|
|
- Permissive prompt warning/fallback behavior
|
|
- Spectrum Interop API v1 with Actual Prefix 2 on continuation chunks
|
|
- Audio Seam Auto
|
|
- Video Seam Auto guarded transient and micro-flash correction
|
|
|
|
Latest verified automated result: 633 passed.
|
|
|
|
Still Image Guide, legacy Main Hi-Res Fix, and Auto 2 remain experimental. Still Image Guide hard-anchor transition quality is not Production-accepted. Main 3x5 2x is not accepted on the tested RTX 5060 Ti 16 GiB because the terminal 77T Second Pass group exceeded VRAM; longer Reference/Hybrid acceptance remains pending.
|
|
|
|
Detailed evidence: docs/VALIDATION_RESULTS.md
|