Squashed 'vendor/dumas_h3_continuum/' content from commit d07ed31

git-subtree-dir: vendor/dumas_h3_continuum
git-subtree-split: d07ed31b442d283d6a1740f875a843b9feda4c91
This commit is contained in:
2026-10-20 13:43:38 +00:00
commit bf8f76505f
481 changed files with 131517 additions and 0 deletions
+33
View File
@@ -0,0 +1,33 @@
# H3 Continuum Join V1 — Manual Wiring
## Clip 1
```text
MiniMax H3 Image to Video ── conditioning/latent ─┐
H3 model → Sage → Sol ───────────────────────────┼→ H3 Continuum Join (state empty)
└→ Spectrum → Basic Guider → Sampler
Sampler output → Video VAE Decode / Audio VAE Decode
└────────────────────────────────────→ H3 Continuum Finish
└→ STATE 1
```
## Clip 2
Duplicate the H3 conditioning/sampling group and connect `STATE 1` to the next `H3 Continuum Join.previous_state`. Use another seed for the next clip.
Two conditioning choices are supported:
1. **Identity anchored:** keep the original image connected. Join removes its spatial first-frame keyframe but retains the Qwen visual tokens as identity guidance.
2. **Context only:** disconnect the image from the second MiniMax H3 Image to Video node. The previous AV state becomes the only visual timeline anchor.
## Clip 3+
Repeat the same pattern. State Save/Load is optional when every clip is in one graph. Use numbered state slots when clips are approved across separate runs.
## Recommended accelerator order
```text
MODEL → SageAttention → Sol-Attn → H3 Continuum Join → Spectrum → Basic Guider
```
The order is shown for clarity. Continuum uses `APPLY_MODEL`, Spectrum uses `DIFFUSION_MODEL`, and Sage/Sol remain attention overrides.
+68
View File
@@ -0,0 +1,68 @@
# H3 Continuum Sampler V2 — 配線
## MODEL
```text
UNET Loader
→ Patch Sage Attention KJ
→ Sol-Attn Patch
→ Power LoRA Loader(任意)
→ Spectrum Apply MiniMax H3
→ H3 Continuum Sampler V2.model
```
## 共有入力
```text
CLIP Loader → clip
Video VAE → video_vae
Audio VAE → audio_vae
KSampler Select → sampler
Basic Scheduler → sigmas
Load Image → ImageScaleToTotalPixels → first_frame
GetImageSize.width/height → width/height(widgetをinputへ変換して接続する場合)
```
## 出力
```text
Sampler V2.images ─→ CreateVideo.images
Sampler V2.audio ─→ CreateVideo.audio
CreateVideo ─→ SaveVideo
Sampler V2.session ─→ Save Session(任意)
```
## 3×5秒の初期値
```text
prompt_mode: Fixed — one prompt
chunks: 3
chunk_seconds: 5.0
continuity: Balanced — 22 frames
base_seed: 固定値
audio_continuity: true
exact_total_duration: true
diagnostics: Basic
reroll_from_chunk: 0
reroll_nonce: 0
strict_compatibility: true
```
## 再開
```text
Load Session.session → Sampler V2.session
chunks: 保存時より大きい値
reroll_from_chunk: 0
```
## 最終チャンクだけリロール
```text
Load Session.session → Sampler V2.session
reroll_from_chunk: 最終チャンク番号
reroll_nonce: 1以上へ変更
```
同梱の`H3_Continuum_V2_3x5s.json`は、上記配線を1本のV2ノードにまとめた編集可能なComfyUI workflow例です。モデル名、入力画像、Prompt、LoRAを自分の環境に合わせて変更してください。
+31
View File
@@ -0,0 +1,31 @@
H3 Continuum V2 standard workflow
Contents:
- H3_Continuum_V2.json
- H3_Continuum_V2_reference.webp
Setup:
1. Copy H3_Continuum_V2_reference.webp to ComfyUI/input/.
2. Drag H3_Continuum_V2.json onto the ComfyUI canvas.
3. Install or replace any custom nodes reported as missing.
4. Confirm the model selections before queueing.
Default profile:
- 3 chunks x 5.0 seconds
- Timeline prompt with three sections
- Scale Image to Total Pixels: area, 0.50 MP, 32-step rounding
- Continuity: Balanced, 22 frames
- Seam correction: Auto
- Report Detail: Basic
Quality note:
- In the current local A/B sample, Seam Correction Off produced the least noticeable visual boundary.
- Compare Auto and Off with the same seed; keep Off when Native Continuity already provides a clean join.
Reference test environment:
- NVIDIA GeForce RTX 5060 Ti 16GB
- 64GB system RAM
- Python 3.13.12, PyTorch 2.13.0+cu130, CUDA 13.0
- SageAttention 2.2.0+cu130, KJNodes FP8 CUDA++, compile disabled
The 0.50 MP profile is a starting point. If VRAM is insufficient, reduce the image scale to 0.30 MP.
Binary file not shown.

After

Width:  |  Height:  |  Size: 47 KiB

File diff suppressed because it is too large Load Diff
File diff suppressed because it is too large Load Diff
File diff suppressed because it is too large Load Diff
@@ -0,0 +1,24 @@
{
"h3_continuum_join": {
"continuity": "Balanced — 22 frames",
"extend_seconds": 5.0,
"audio_continuity": true,
"first_frame_policy": "Use Continuum context; keep visual identity tokens (Recommended)",
"preserve_last_frame": true,
"strict_compatibility": true,
"debug": false
},
"spectrum": {
"blend_weight": 0.5,
"degree": 1,
"warmup_steps": 1,
"offline_smoothing_replay": true,
"audio_blend_weight": 0.0
},
"validation_matrix": [
"Sage only",
"Sage + Sol",
"Sage + Spectrum",
"Sage + Sol + Spectrum"
]
}
@@ -0,0 +1,29 @@
{
"h3_continuum_sampler_v2": {
"chunks": 3,
"chunk_seconds": 5.0,
"continuity": "Balanced — 22 frames",
"audio_continuity": true,
"exact_total_duration": true,
"diagnostics": "Basic",
"reroll_from_chunk": 0,
"reroll_nonce": 0,
"strict_compatibility": true,
"debug": false
},
"sampling": {
"sampler": "res_multistep",
"scheduler": "simple",
"steps": 20,
"denoise": 1.0,
"turbo_lora": false
},
"spectrum": {
"enabled": true,
"blend_weight": 0.5,
"degree": 1,
"warmup_steps": 1,
"offline_smoothing_replay": true,
"audio_blend_weight": 0.0
}
}
+5
View File
@@ -0,0 +1,5 @@
The same woman begins singing in one continuous shot. Preserve identity, clothing, background, lighting, camera direction, vocal timbre, and piano tempo. End while the phrase is still continuing.
---
Seamless continuation of the immediately preceding clip. Continue the exact mouth motion, sustained note, breathing state, camera movement, vocal key, tempo, and piano phrase. No cut or reset.
---
Seamless final continuation. Preserve all identity and motion. Complete the vocal phrase and piano cadence naturally during the final second, then release a soft breath. No scene change.
+8
View File
@@ -0,0 +1,8 @@
[0-5s]
The same woman begins singing in a single continuous shot. The melody and camera movement remain unresolved at the boundary.
[5-10s]
Continue the exact performance, mouth motion, breath, vocal key, piano tempo, and camera movement with no reset.
[10-15s]
Continue seamlessly, then resolve the final vocal note and piano phrase naturally in the last second.
File diff suppressed because it is too large Load Diff
File diff suppressed because it is too large Load Diff
Binary file not shown.
File diff suppressed because it is too large Load Diff
File diff suppressed because it is too large Load Diff
File diff suppressed because it is too large Load Diff
Binary file not shown.
File diff suppressed because it is too large Load Diff
Binary file not shown.
File diff suppressed because it is too large Load Diff
File diff suppressed because it is too large Load Diff