Files
chris.dumas 3716b02a57
tests / test (push) Canceled after 0s
tests / windows-storage (push) Canceled after 0s
Archive fork in favor of upstream and DUMAS nodes
2026-10-20 10:46:16 +00:00

70 lines
4.6 KiB
Markdown
Raw Permalink Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# 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。