This repository provides FlashPortrait custom nodes for ComfyUI. It allows you to generate infinite-length portrait animations driven by a video, directly within your ComfyUI workflow.

Original Paper: FlashPortrait: 6x Faster Infinite Portrait Animation with Adaptive Latent Prediction

• Automatic Model Downloading: No manual weight placement required. The loader fetches Wan2.1 and FlashPortrait models from Hugging Face automatically.
• Infinite Length Support: Uses FlashPortrait's sliding window mechanism to process long driving videos.
• Flexible Inputs:
  • Reference Image: Defines the identity (ID).
  • Driving Video: Defines the motion and expression.

1. Clone the Repository: Go to your ComfyUI custom_nodes folder and run:

   cd ComfyUI/custom_nodes
   git clone https://github.com/okdalto/ComfyUI-FlashPortrait

2. Install Dependencies:

   cd ComfyUI/custom_nodes/ComfyUI-FlashPortrait
   pip install -r requirements.txt

   Note: Requires approx. 40GB VRAM for full model loading (BF16).

3. Restart ComfyUI.

You can find a basic ComfyUI workflow in examples/flash_portraits_example.json. Simply drag and drop this file into your ComfyUI window to load the graph.

Loads the heavy models (Transformer, VAE) and face alignment tools.

• precision: bf16 (recommended), fp16, fp32.
• GPU_memory_mode: default, sequential_cpu_offload, model_cpu_offload_and_qfloat8.
• download_missing: Enable this to automatically download models to ComfyUI/models/flash_portrait/ on first run.

Extracts motion and expression features from the driving video.

• images: Connect your driving video frames here (use Load Video or similar nodes).
• source_fps: Frame rate of the source video (default 25.0).
• context_size / context_overlap: Controls the sliding window size for feature extraction.

The main generation node.

• pipe: Connect from Loader.

• head_emo_features: Connect from Feature Extractor.

• image: The Reference Image (Identity). Only the first frame is used.

• prompt / negative_prompt: Text guidance.

• guidance_scale / text_cfg_scale / emo_cfg_scale: Control the influence of unconditional, text, and emotion guidance.

• steps: Inference steps (default 30).

• max_size: Output resolution height (e.g., 720).

This implementation is based on the amazing work by the original FlashPortrait team. Huge congratulations and thanks to: Shuyuan Tu, Yueming Pan, Yinming Huang, Xintong Han, Zhen Xing, Qi Dai, Kai Qiu, Chong Luo, Zuxuan Wu

And to the open-source projects that made this possible:

• Wan2.1
• PD-FGC
• FantasyPortrait
• VideoX-Fun

@article{tu2025flashportrait,
  title={FlashPortrait: 6x Faster Infinite Portrait Animation with Adaptive Latent Prediction},
  author={Tu, Shuyuan and Pan, Yueming and Huang, Yinming and Han, Xintong and Xing, Zhen and Dai, Qi and Qiu, Kai and Luo, Chong and Wu, Zuxuan},
  journal={arXiv preprint arXiv:2512.16900},
  year={2025}
}