ComfyUI SynVow TeleStyle

A ComfyUI video style transfer plugin based on TeleStyle.

Original Project

https://github.com/Tele-AI/TeleStyle

Features

• 🎬 Video Style Transfer: Apply the style from a reference image to source video
• 🤖 Auto Model Download: Automatically download required models from HuggingFace on first use
• ⚙️ Flexible Parameter Control: Adjust resolution, frame count, inference steps, guidance scale, etc.
• 📊 Real-time Progress Display: Show progress bar during inference
• 🎯 VRAM Optimization: Built-in VAE Tiling to reduce VRAM usage

Installation

1. Install Dependencies

pip install -r requirements.txt

Or manually install core dependencies:

pip install transformers diffusers pillow decord opencv-python safetensors omegaconf einops sentencepiece accelerate peft ftfy regex

2. Model Download

Models will be automatically downloaded to the following directory on first use:

ComfyUI/models/TeleStyle/
├── Wan2.1-T2V-1.3B-Diffusers/    # Wan2.1 base model
├── dit.ckpt                       # TeleStyle style transfer weights
└── prompt_embeds.pth              # Pre-computed text embeddings

Supported base models:

• Wan2.1-T2V-1.3B - 1.3B parameters, lower VRAM requirement
• Wan2.1-T2V-14B - 14B parameters, better quality but higher VRAM requirement

Node Description

1. TeleStyle Video Loader

Load TeleStyle video style transfer model.

Input Parameters:

• wan_model: Select Wan2.1 base model (1.3B or 14B)

Output:

• model: Model info for subsequent style transfer node

2. TeleStyle Video

Perform video style transfer.

Input Parameters:

• model: Model from Loader node
• video_frames: Source video frames (IMAGE type)
• style_first_frame: Style reference image (typically a stylized first frame)
• height: Output video height (default 1024)
• width: Output video width (default 1024)
• video_length: Output video frame count (default 81)
• num_inference_steps: Inference steps (default 25, fewer = faster but lower quality)
• guidance_scale: Guidance strength (default 3.0, higher = stronger style transfer)
• seed: Random seed

Output:

• video_frames: Style-transferred video frames (IMAGE type)

Usage Example

Basic Workflow

1. Use Load Video node to load source video
2. Use Load Image node to load style reference image (stylized first frame)
3. Connect TeleStyle Video Loader to load model
4. Connect TeleStyle Video for style transfer
5. Use Video Combine node to save output video

Recommended Parameters (24GB VRAM)

• height: 480
• width: 832
• video_length: 49
• num_inference_steps: 20-25

Notes

1. VRAM Requirements:

   • 1.3B model: 16GB+ VRAM recommended
   • 14B model: 40GB+ VRAM recommended
   • Reduce resolution and frame count to lower VRAM usage

2. Style Reference Image:

   • Typically use image style transfer tools (e.g., TeleStyle Image) to stylize the first frame
   • The style reference image affects the entire video's style

3. Inference Speed:

   • Fewer inference steps = faster but lower quality
   • Recommend testing with fewer steps (e.g., 10-15) first

Troubleshooting

CUDA Out of Memory

• Lower height and width parameters
• Reduce video_length frame count
• Use 1.3B model instead of 14B

Output Video is Black

• Check if style reference image is correctly connected
• Ensure source video frames are in correct format (IMAGE type, value range 0-1)

License

This project follows the license of the original TeleStyle project.

Related Links

• TeleStyle Original Project
• Wan2.1 Models