A custom node for ComfyUI that enables coherent video generation while maintaining efficient memory usage, specifically optimized for heavy models like Flux.

Features

🎥 Frame-by-frame video processing with motion preservation
🧠 Efficient memory management for heavy models
🔄 Progressive denoising with coherence maintenance
💫 Dynamic quality control and motion guidance
🎨 Style preservation across frames

Installation

Navigate to your ComfyUI custom nodes directory:

cd ComfyUI/custom_nodes

Clone this repository:

git clone https://github.com/ShmuelRonen/ComfyUI-CohernetVideoSampler.git

Restart ComfyUI

Usage

The node appears in the node menu as "Cohernet Video Sampler".

Inputs

model: Your diffusion model (tested extensively with Flux)
positive: Positive prompt conditioning
negative: Negative prompt conditioning
video_latents: Input video in latent space (from VAE Encode)
seed: Generation seed
steps: Number of sampling steps
cfg: Classifier free guidance scale
sampler_name: Choice of sampler
scheduler: Choice of scheduler
denoise: Denoising strength
motion_strength: Strength of motion preservation

Memory Management

The node implements several memory optimization techniques:

Progressive batch processing
Automatic VRAM cleanup
Dynamic batch size adjustment
Efficient latent space operations

This allows it to work smoothly even with memory-intensive models like Flux without OOM errors.

Example Workflow

Load your video
Encode it to latent space using VAE Encode
Connect to Coherent Video Sampler
Set your sampling parameters
Generate coherent video output

Parameters Guide

denoise: Lower values (0.4-0.6) preserve more of the original motion
motion_strength: Higher values (0.7-0.9) maintain stronger frame-to-frame coherence
steps: Even with heavy models, 20-30 steps usually sufficient
cfg: Standard values (7-9) work well for most cases

Memory Usage Examples

When using with Flux model:

20 frame video @ 512x512: ~8GB VRAM
40 frame video @ 512x512: ~10GB VRAM
Processing happens in windows of frames to maintain stable memory usage

Known Limitations

Very long videos might need to be processed in segments
Extreme motion can affect coherence
High denoise values might reduce motion preservation

Future Plans

Additional motion control parameters
Custom denoising patterns
Advanced style preservation options
Multi-model support optimization

Contributing

Pull requests are welcome. For major changes, please open an issue first to discuss what you would like to change.

License

MIT License

Acknowledgments

ComfyUI team for the amazing framework
Flux model team for the inspiration in handling heavy models