ComfyUI-LatentSyncWrapper 1.5

Support My Work

If you find this project helpful, consider buying me a coffee:

Unofficial LatentSync 1.5 implementation for ComfyUI on Windows and WSL 2.0.

This node provides advanced lip-sync capabilities in ComfyUI using ByteDance's LatentSync 1.5 model. It allows you to synchronize video lips with audio input with improved temporal consistency and better performance on a wider range of languages.

What's new in LatentSync 1.5?

1. Temporal Layer Improvements: Corrected implementation now provides significantly improved temporal consistency compared to version 1.0
2. Better Chinese Language Support: Performance on Chinese videos is now substantially improved through additional training data
3. Reduced VRAM Requirements: Now only requires 20GB VRAM (can run on RTX 3090) through various optimizations:
   • Gradient checkpointing in U-Net, VAE, SyncNet and VideoMAE
   • Native PyTorch FlashAttention-2 implementation (no xFormers dependency)
   • More efficient CUDA cache management
   • Focused training of temporal and audio cross-attention layers only
4. Code Optimizations:
   • Removed dependencies on xFormers and Triton
   • Upgraded to diffusers 0.32.2

Prerequisites

Before installing this node, you must install the following in order:

1. ComfyUI installed and working

2. FFmpeg installed on your system:

   • Windows: Download from here and add to system PATH

Installation

Only proceed with installation after confirming all prerequisites are installed and working.

1. Clone this repository into your ComfyUI custom_nodes directory:

cd ComfyUI/custom_nodes
git clone https://github.com/ShmuelRonen/ComfyUI-LatentSyncWrapper.git
cd ComfyUI-LatentSyncWrapper
pip install -r requirements.txt

Required Dependencies

diffusers>=0.32.2
transformers
huggingface-hub
omegaconf
einops
opencv-python
mediapipe
face-alignment
decord
ffmpeg-python
safetensors
soundfile

Note on Model Downloads

On first use, the node will automatically download required model files from HuggingFace:

• LatentSync 1.5 UNet model
• Whisper model for audio processing
• You can also manually download the models from HuggingFace repo: https://huggingface.co/ByteDance/LatentSync-1.5

Checkpoint Directory Structure

After successful installation and model download, your checkpoint directory structure should look like this:

./checkpoints/
|-- .cache/
|-- auxiliary/
|-- whisper/
|   `-- tiny.pt
|-- config.json
|-- latentsync_unet.pt  (~5GB)
|-- stable_syncnet.pt   (~1.6GB)

Make sure all these files are present for proper functionality. The main model files are:

• latentsync_unet.pt: The primary LatentSync 1.5 model
• stable_syncnet.pt: The SyncNet model for lip-sync supervision
• whisper/tiny.pt: The Whisper model for audio processing

Usage

1. Select an input video file with AceNodes video loader
2. Load an audio file using ComfyUI audio loader
3. (Optional) Set a seed value for reproducible results
4. (Optional) Adjust the lips_expression parameter to control lip movement intensity
5. (Optional) Modify the inference_steps parameter to balance quality and speed
6. Connect to the LatentSync1.5 node
7. Run the workflow

The processed video will be saved in ComfyUI's output directory.

Node Parameters:

• video_path: Path to input video file
• audio: Audio input from AceNodes audio loader
• seed: Random seed for reproducible results (default: 1247)
• lips_expression: Controls the expressiveness of lip movements (default: 1.5)
  • Higher values (2.0-3.0): More pronounced lip movements, better for expressive speech
  • Lower values (1.0-1.5): Subtler lip movements, better for calm speech
  • This parameter affects the model's guidance scale, balancing between natural movement and lip sync accuracy
• inference_steps: Number of denoising steps during inference (default: 20)
  • Higher values (30-50): Better quality results but slower processing
  • Lower values (10-15): Faster processing but potentially lower quality
  • The default of 20 usually provides a good balance between quality and speed

Tips for Better Results:

• For speeches or presentations where clear lip movements are important, try increasing the lips_expression value to 2.0-2.5
• For casual conversations, the default value of 1.5 usually works well
• If lip movements appear unnatural or exaggerated, try lowering the lips_expression value
• Different values may work better for different languages and speech patterns
• If you need higher quality results and have time to wait, increase inference_steps to 30-50
• For quicker previews or less critical applications, reduce inference_steps to 10-15

Known Limitations

• Works best with clear, frontal face videos
• Currently does not support anime/cartoon faces
• Video should be at 25 FPS (will be automatically converted)
• Face should be visible throughout the video

Credits

This is an unofficial implementation based on:

• LatentSync 1.5 by ByteDance Research
• ComfyUI

License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.