https://github.com/fblissjr/ComfyUI-WanSeamlessFlow
experimental wanvideo comfyui node with a singular goal - visually seamless transitions between context windows
https://github.com/fblissjr/ComfyUI-WanSeamlessFlow
Last synced: 3 months ago
JSON representation
experimental wanvideo comfyui node with a singular goal - visually seamless transitions between context windows
- Host: GitHub
- URL: https://github.com/fblissjr/ComfyUI-WanSeamlessFlow
- Owner: fblissjr
- License: apache-2.0
- Created: 2025-03-16T14:22:20.000Z (3 months ago)
- Default Branch: main
- Last Pushed: 2025-03-16T15:12:48.000Z (3 months ago)
- Last Synced: 2025-03-16T15:27:31.621Z (3 months ago)
- Language: Python
- Size: 13.7 KB
- Stars: 0
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
- awesome-comfyui - **wanvideo - seamless flow** - visually seamless transitions between context windows (All Workflows Sorted by GitHub Stars)
README
# wanvideo - seamless flow
```
ComfyUI-WanSeamlessFlow/
├── __init__.py # Registry and imports
├── blending.py # Core embedding interpolation functions
├── nodes.py # ComfyUI node definitions
├── visualization.py # Diagnostic visualization utilities
├── README.md # Documentation and examples
└── utils/ # Support utilities
└── optimization.py # Embedding optimization algorithms
```
## key notes - needs modifications, for now, to Kijai's wanvideo wrapper
- see `./reference/nodes.py` for current patches made:
## architecture / data flow map
```
[Architecture Map]
┌─────────────────────┐ ┌───────────────────────┐ ┌─────────────────────┐
│ WanSeamlessFlow │ ──→ │ Context Window Engine │ ──→ │ Rendering Pipeline │
│ • Embedding Order │ │ • Window Transition │ │ • Composite Output │
│ • Blend Parameters │ │ • Interpolation │ │ • Visual Smoothing │
└─────────────────────┘ └───────────────────────┘ └─────────────────────┘
```
## integration with [Kijai's ComfyUI-WanVideoWrapper](https://github.com/kijai/ComfyUI-WanVideoWrapper)
```
┌──────────────────┐ ┌──────────────────┐ ┌──────────────────┐ ┌──────────────────┐
│ LoadWanVideo │ │ WanVideoText │ │ WanSmartBlend │ │ WanVideoSampler │
│ T5TextEncoder │───▶│ Encode │───▶│ │───▶│ │
└──────────────────┘ └──────────────────┘ └──────────────────┘ └──────────────────┘
│
▼
┌──────────────────┐
│ WanBlendVisualize│
│ (Optional) │
└──────────────────┘
```
### usage
**Multi-Prompt Usage**:
For optimal results with prompt transitions:
Modify your WanVideoTextEncode to use multiple prompts separated by | characters:
`high quality nature video featuring a red panda balancing on a bamboo stem | high quality nature video focusing on the bird perched on the panda's head | high quality nature video showcasing the waterfall in the background`
- Adjust the blend_width parameter based on your number of frames:
- With 257 frames and 3 prompts → 85.6 frames per prompt
- Recommended blend_width: 8-16 frames
- Higher values create wider transition zones
**Compatibility Notes**:
This setup is fully compatible with your existing components:
- TeaCache: Works alongside WanSmartBlend, both optimizing different parts
- Context Windowing: Seamless transitions work at context window boundaries
- Torch Compilation: No interference, remains performance-enhancing
**Parameter Recommendations**:
- for your particular setup with 257 frames:
- blend_width: 8 # Start conservative, increase for smoother transitions
- blend_method: "smooth" # Provides natural transitions without obvious linear interpolation
- optimize_order: true # Automatically orders prompts for minimal semantic distance
- verbosity: 1 # Basic logging without overwhelming console output
**Extended Analysis**:
This integration creates a multi-dimensional benefits matrix:
```
⎡ TeaCache Compatibility ⎤ ⎡ High | Compatible with caching mechanisms ⎤
⎢ Context Window Flow ⎥ = ⎢ High | Works with all scheduler types ⎥
⎢ Smooth Transitions ⎥ ⎢ High | Creates gradual prompt blending ⎥
⎢ Performance Impact ⎥ ⎢ Low | Minimal computational overhead ⎥
⎣ Implementation Effort ⎦ ⎣ Low | Non-invasive integration ⎦
```
## logical flow
**Integration point: context window embedding selection logic**
```
WindowProcessingPipeline {
window_context → embedding_selection → model_forward → window_composition
↑ ↑ ↑
| (context info) | (embedding selection) | (output compositing)
↓ ↓ ↓
context_scheduler [INTERVENTION POINT] window_blending
}
```