ComfyUI-TeleStyle

An unofficial, streamlined, and highly optimized ComfyUI implementation of TeleStyle.

This node is specifically designed for Video Style Transfer using the Wan2.1-T2V architecture and TeleStyle custom weights. Unlike the original repository, this implementation strips away all heavy image-editing components (Qwen weights) to focus purely on video generation with speed/quality.

Requirements

• GPU VRAM: 6GB minimum
• Disk Space: ~6GB for models and weights

✨ Key Features

• High Performance:

  • Acceleration: Built-in support for Flash Attention 2 and SageAttention for faster inference.
  • Fast Mode: Optimized memory management with aggressive cache cleanup to prevent conflicts between CPU offloading and GPU processing.

• Simplified Workflow: No need for complex external text encoding nodes. The model uses pre-computed stylistic embeddings (prompt_embeds.pth) for maximum efficiency.

h-7.mp4

h-6.mp4

w-9.mp4

h-8.mp4

📦 Installation

Navigate to your ComfyUI custom nodes directory:

cd ComfyUI/custom_nodes/

Clone this repository:

git clone https://github.com/neurodanzelus-cmd/ComfyUI-TeleStyle.git

Install dependencies:

pip install -r requirements.txt

Note: For SageAttention support, you may need to install sageattention manually.

📂 Model Setup

This node requires specific weights placed in the ComfyUI/models/telestyle_models/ directory.

The weights are downloaded automatically at the first run

Directory Structure:

ComfyUI/
└── models/
    └── telestyle_models/
        ├── weights/
        │   ├── dit.ckpt            # Main Video Transformer weights
        │   └── prompt_embeds.pth   # Pre-computed style embeddings
        └── Wan2.1-T2V-1.3B-Diffusers/
        │   ├── transformer_config.json
        │   ├── vae/
        │   │   │   ├── diffusion_pytorch_model.safetensors
        │   │   │   └── config.json
        │   └── scheduler/
        │       └── scheduler_config.json

Where to get weights:

https://huggingface.co/Danzelus/TeleStyle_comfy/tree/main

🚀 Usage

1. TeleStyle Model Loader

This node loads the necessary model components.

Parameter	Description
dtype	Choose between bf16 (best quality), fp16

2. TeleStyle Video Transfer

The main inference node.

Parameter	Description
model	Connect the output from the Loader
video_frames	Input video batch (from Load Video or VHS_LoadVideo)
style_image	A reference image to guide the style transfer
steps	Inference steps (default: 12)
cfg	Guidance scale (default: 1)
scheduler	Choose your sampler (FlowMatchEuler, DPM++)
fast_mode	Keep True for speed. Set to False for low-VRAM offloading (slower)
acceleration	default - Standard PyTorch attention flash_attn - Faster, requires compatible GPU sage_attn - Ultra-fast, requires sageattention library

---

To-Do List

• Initial release
• More samplers
• Consistency for very long videos

---

Guys, I’d really appreciate any support right now. I’m in a tough spot:

📜 Credits

This project is an unofficial implementation based on the amazing work by the original authors. Please refer to their repository for the original research and model weights.