docs: explain partial regeneration and refresh node images

This commit is contained in:
ukr8b3g-cmyk
2026-08-28 12:08:24 +09:00
parent 8dec0b7ceb
commit ecc72dae9b
6 changed files with 70 additions and 5 deletions
+4 -2
View File
@@ -24,12 +24,14 @@ c9573d4393236276d032c4503f8ed8bd19c6ef24e84eec6f64f8a4c47aabc718 docs/images/v3
ff8d901d05aa89f9b698c60b3d11534ef97982b3e22e9ede6255a85149a669c4 docs/images/v34-video-reference-separate-audio.png
06c333b3dda41526da70d56f4aa25922aafa2f48ec2d99350b0ed6b428e92bad docs/images/v34-video-reference-with-audio.png
301afedc3e0922a4bf3bcf445734b5adcdd5c35384132b24a38098c66750f499 docs/images/v34-workflow-overview.png
4e0d904d3bad89cb8c3ffe17fb645bfb6fbc87fe49fe3c75c91094b983726025 docs/images/v35-assemble-seam-node.png
a53796f7a904d675e80a2276c20d68502f03a3958529278e6ce9465baaf1c4af docs/images/v35-hires-fix-node.png
5a23f232a8b795226e8255e6e2e3248b0130f1c7ce276d50370d0333d6a62ef0 docs/images/v35-second-pass-node.png
04a34b1b55b627553d08b7b10de71295d55ea9e4ac9815b6dd8ce6d268098851 docs/images/v351-lbh-conditioning-bridge-flow.svg
b16d63242b45fe1abef0a4dddd4fc5b9306aa30c2f84bbc3ddf334a1c0832957 docs/images/v351-reference-audio-permanent.png
ed189c7f16ba1f6e48f557312066fdb5e1313b5ea7df853a99467fd0d7532ef5 docs/images/v351-video-guide-frames.png
0c1dfd5acff349b98ecdfb57ae50941f6da26a462dfd2aaa9b7c7b0a34abf304 docs/images/v352-stabilization-optimization.png
925898dfe741fdbcc9b429133cfcbfb8ff718a3b476b368a063b46d39511e2f9 docs/images/v36-sampler-node.png
93f07aa727b34802bce3d333426b8ba2af522500882a6d222d00cf577ca82e49 docs/SEAM_TEST_SOURCE_JA.md
2c13aefd137db92a3dbf5576e9442d58012b1f0de1240c07d039ceeb702e7c1f docs/V2_1_4_UI_SIMPLIFICATION.md
b08022df068576c759c8bf80ac6b6519693f3c65c917f1822aeb06ea359aa5b1 docs/V2_1_5_UI_SERIALIZATION_STABILITY.md
@@ -75,8 +77,8 @@ f456a678af61f0feeae009bce89db763b60c820444d75ff5d2a30bb74991c66b PACKAGE_INFO_J
b470819f16fc179654dc1600721a63f0a5a14e777f4c004d574b499da6886ba5 public_api.py
97415d587e8ab835f1c22a375fe64d73541440f86f11afeace15fed69c1ae839 pyproject.toml
94fdc2d545b2d4077a4d70229fd22ba3ccac21d49d1097d2873f76d04e9a3c7a pytest.ini
281b541e53fa9aaede3a141f5d24bde8ef80edb18169326d21275cd83709a7b2 README_JA.md
1103cda983da3278ec056282c596804bb05200c276ccf0cfa2b1b6170a07741c README.md
c9000b17f8cb5b6e19bb08c74b9aba26166e6a4950b11da52b0c4c951fe4b14e README_JA.md
ad802d64cf8fa4c4d35c349e30c56de72792cb06337033f990d282799717ac8c README.md
5d89c366b4f076fa73a2695557e0af749bf95df1dad0d55dbb2fe5b1c2519ad7 reference_audio.py
a0266ab4434b8e9844219d2b6dd8a168d198dd8f142775be01a39b81dc196b91 reference_video.py
564f5f2bf29665059845a98b9ca3e9a1309566aab73bcaa96f95fab8251ab06c reference.py
+26 -2
View File
@@ -339,9 +339,11 @@ Hard stops remain only for states that cannot execute safely, including corrupt
### Cleaner interface
The public nodes focus on normal production controls. Developer diagnostics and detailed reports are available through ComfyUI settings.
The public nodes focus on normal production controls. Developer diagnostics and detailed reports are available through ComfyUI settings. The current production path uses the V3.6 sampler with the V3.5 low-memory assembler.
![V3.4 sampler, Spectrum, and assembler](docs/images/v34-sampler-spectrum-assemble.png)
![H3 Continuum Sampler V3.6](docs/images/v36-sampler-node.png)
![H3 Continuum Assemble + Seam V3.5](docs/images/v35-assemble-seam-node.png)
### Timeline paths
@@ -568,6 +570,28 @@ Run Storage preserves completed chunks and can reuse them after restarting Comfy
![Regenerate From](docs/images/v34-regenerate-from.png)
### Keep the first part and regenerate the second part
For a normal T2VA or I2VA run configured as `2 chunks x 5 seconds`, Chunk 1 is the first five seconds and Chunk 2 is the second five seconds.
Before the first generation:
1. Set `Run Storage` to `Save + Auto Resume`.
2. Choose a fixed `Run Name` and keep the Base Seed fixed.
3. Generate the complete ten-second video once. Continuum stores the raw Video/Audio latent for each completed chunk.
To generate a different second part while keeping the first part:
1. Keep the same Run Name, Base Seed, model, LoRA, resolution, sampler, SIGMAS/steps, references, and Continuation settings.
2. Set `Regenerate From` to `Chunk 2`.
3. Queue the workflow again.
Continuum loads the stored Chunk 1 instead of Sampling it again, uses its finalized end as the continuation context, and Samples a new Chunk 2. The report should show `1 reused, 1 generated`. The complete output is then decoded and assembled again; the saved first chunk is reused, while boundary Seam processing is evaluated during the new assembly.
If the two parts need different instructions, define them with List or Timeline Prompt syntax before the first generation. Changing the Prompt Plan or other generation-contract inputs after a stored run can create a new revision, in which case the previous Chunk 1 may not be reusable. A run originally generated with `Run Storage = Off` cannot be reused retroactively. `Regenerate From` works at chunk boundaries; it does not edit an arbitrary region inside one chunk.
FL2VA Long Terminal Merge is an important exception: its final two logical chunks are one atomic physical Sampling unit, so those two chunks are regenerated together.
- Use a fixed seed for reproducible resume.
- Changing persistent models, LoRAs, references, prompts, resolution, or sampling settings can create a new revision.
- Unknown upstream node classes no longer cause rejection by themselves; reuse still depends on compatible observable contracts and hashes.
+27 -1
View File
@@ -187,7 +187,11 @@ Hi-Res Fixを接続しなければV3.4 Sampling経路は変わりません。強
- [V3.5互換テンプレート](examples/workflows/MiniMax_H3_Continuum_V35.json)
- [V3.5.1 LBH+Conditioning Bridge接続例](examples/workflows/MiniMax_H3_Continuum_V351_LBH_Conditioning_Bridge.json)
V3.6テンプレートはContinuation BackendがStandard、Hi-Res Fixが初期OFF、AssembleのBuffer BackendがAutoです。First Frame、Last Frame、Reference Image 1~3は1つのmegapixel設定を共有します。`Video Guide Size`は`Video Guide Frames`に使うframe batchを調整し、decodeとframe-rate変換はVideo loader側で行います。V3.5テンプレートは互換用として変更せず維持します。
V3.6テンプレートはContinuation BackendがStandard、Hi-Res Fixが初期OFF、AssembleのBuffer BackendがAutoです。First Frame、Last Frame、Reference Image 1~3は1つのmegapixel設定を共有します。`Video Guide Size`は`Video Guide Frames`に使うframe batchを調整し、decodeとframe-rate変換はVideo loader側で行います。現在のProduction構成はV3.6 SamplerとV3.5の低メモリAssembleを組み合わせます。V3.5テンプレートは互換用として変更せず維持します。
![H3 Continuum Sampler V3.6](docs/images/v36-sampler-node.png)
![H3 Continuum Assemble + Seam V3.5](docs/images/v35-assemble-seam-node.png)
## V3.4互換
@@ -210,6 +214,28 @@ V3.5のCompact UIでは、対応する入力を接続した時だけSize項目
Reference入力のバイパスされたソケットは無視され、有効な画像だけがPicture 1から連続採番されます。空欄や不完全なプロンプトも停止せず、警告とFixed fallbackで生成を続行します。未知のモデル、ノード、プロンプト形式を独自判断で拒否する制限も撤廃しました。
## Run Storage:第2部分だけを再生成
通常のT2VA/I2VAを`2 chunks x 5 seconds`にした場合、Chunk 1が最初の5秒、Chunk 2が次の5秒です。
初回生成前に次を設定します。
1. `Run Storage`を`Save + Auto Resume`にします。
2. 固定の`Run Name`を選び、Base Seedも固定します。
3. まず10秒全体を一度生成します。Continuumは完了した各Chunkのraw Video/Audio latentを保存します。
最初の5秒を保持して、第2部分だけを作り直す手順は次のとおりです。
1. Run Name、Base Seed、model、LoRA、解像度、sampler、SIGMAS/steps、Reference、Continuity設定を初回と同じに保ちます。
2. `Regenerate From`を`Chunk 2`にします。
3. Workflowをもう一度Queueします。
Continuumは保存済みChunk 1を再Samplingせずに読み込み、その確定済み末尾をContinuation contextとして新しいChunk 2をSamplingします。Reportの目安は`1 reused, 1 generated`です。その後、10秒全体を再Decode/Assemblyします。第1ChunkのSampling結果は再利用され、境界のSeam処理は新しいAssembly時に再評価されます。
第1部分と第2部分で異なる指示を使う場合は、初回生成前にListまたはTimeline Promptとして分けておきます。保存後にPrompt Planや生成Contractの入力を変更すると新しいRevisionになり、以前のChunk 1を再利用できない場合があります。初回を`Run Storage = Off`で生成した場合も、後から遡って再利用することはできません。`Regenerate From`はChunk境界単位であり、1つのChunk内部の任意領域を編集する機能ではありません。
FL2VA Long Terminal Mergeは重要な例外です。最後の2つのlogical Chunkを1つのatomicなphysical Samplingとして扱うため、その2つは一緒に再生成されます。
Run Storageの保存契約にはDriving AudioとVideo Guide Framesのidentityを含め、入力変更後に古いChunkを誤再利用しないようにしています。
## Driving Audio / Video Guide Frames / Seam
+13
View File
@@ -1,5 +1,18 @@
# Work Log
## 2026-08-28 - Current V3.6 sampler README images
- Replaced the English README's obsolete V3.4 combined workflow screenshot with separate current `H3 Continuum Sampler V3.6` and `H3 Continuum Assemble + Seam V3.5` node images, and added the same images to the Japanese recommended-template section.
- Clarified that the current Production path combines the V3.6 sampler with the V3.5 low-memory assembler. No node, workflow, schema, or runtime behavior changed.
- Snapshot: `pre-readme-v36-node-images-20260828_121500-8dec0b7` (219 tracked files, zero missing or SHA-256 mismatches, revision `8dec0b7ceb961102e36de4dfa735eb40c4cfa4c1`).
- Validation: all 180 Manifest entries PASS, all 13 unique README image links resolve, focused README/release/workflow tests `11 passed`, and `git diff --check` PASS.
## 2026-08-28 - Run Storage partial-regeneration README guide
- Expanded the English and Japanese Run Storage documentation with a concrete normal T2VA/I2VA `2 chunks x 5 seconds` procedure: enable `Save + Auto Resume` before the first run, retain the Run Name and generation contract, select `Regenerate From = Chunk 2`, and confirm `1 reused, 1 generated`.
- Documented that Prompt Plan or other contract changes can create a new revision, `Run Storage = Off` cannot be reused retroactively, regeneration operates at chunk boundaries, reassembly reevaluates Seam processing, and FL2VA Long Terminal Merge regenerates its final atomic pair together. No Production code, UI, workflow, or schema changed.
- Snapshot: `pre-readme-run-storage-regenerate-guide-20260828_115840-8dec0b7` (630 files, mismatch 0, revision `8dec0b7ceb961102e36de4dfa735eb40c4cfa4c1`). Validation: 178 Manifest entries PASS, focused README/release/workflow tests `11 passed`, and `git diff --check` PASS.
## 2026-08-28 - README ComfyUI compatibility requirement
- Added a concise English/Japanese installation note: ComfyUI `>=0.32.0` is required, package validation passed on `0.33.3`, V3.6.1 GPU Smoke passed on `0.34.0`, and the latest ComfyUI release is not mandatory.
Binary file not shown.

After

Width:  |  Height:  |  Size: 26 KiB

Binary file not shown.

After

Width:  |  Height:  |  Size: 72 KiB