This is a ComfyUI custom node implementation of PersonaLive: Expressive Portrait Image Animation for Live Streaming.

Original Repository: GVCLab/PersonaLive
Paper: ArXiv 2512.11253

I deeply appreciate the authors Zhiyuan Li, Chi-Man Pun, Chen Fang, Jue Wang, and Xiaodong Cun for their amazing work and for sharing their code.

1. Clone this repository into your ComfyUI/custom_nodes/ directory:

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

2. Model Setup:

   Option 1: Automatic Download (Recommended)

   Models will be automatically downloaded from Hugging Face when you first use the PersonaLiveCheckpointLoader node. The node will:

   • Detect missing models in your selected directory
   • Download required models (~15-20GB total):
     • lambdalabs/sd-image-variations-diffusers (Base Model)
     • stabilityai/sd-vae-ft-mse (VAE)
     • huaichang/PersonaLive (PersonaLive Weights)
   • Organize them into the correct structure automatically

   [!NOTE] The first download will take some time depending on your internet speed. Models are cached locally, so subsequent loads will be instant.

   Option 2: Manual Download

   If you prefer to download models manually or have connectivity issues:

   Create a folder named persona_live inside your ComfyUI/models/ directory with the following structure:

   ComfyUI/models/
   └── persona_live/
       ├── sd-image-variations-diffusers/  <-- Base Model
       ├── sd-vae-ft-mse/                  <-- VAE
       └── persona_live/                   <-- PersonaLive Repository
           └── pretrained_weights/         <-- .pth files location
               └── personalive/            <-- .pth files location    
                   ├── denoising_unet.pth
                   ├── motion_encoder.pth
                   ├── motion_extractor.pth
                   ├── pose_guider.pth
                   ├── reference_unet.pth
                   └── temporal_module.pth
   

   • Download Base Models:
     • sd-image-variations-diffusers
     • sd-vae-ft-mse
   • Download PersonaLive Weights:
     • Clone or download the entire repository from Hugging Face

1. PersonaLiveCheckpointLoader: Select the model_dir (e.g., persona_live) that contains all your models.
2. PersonaLivePhotoSampler:
   • Connect the pipeline from the loader.
   • Connect ref_image (source portrait) and driving_image (pose reference). driving_image can be a single image or a video frame batch.
   • Set width and height (default 512). The node automatically resizes inputs to this resolution for processing and then restores the original resolution for the output.

• Input Images: It is highly recommended to use square images (1:1 aspect ratio) for both ref_image and driving_image to ensure the best face alignment and generation quality.
• Inference Steps: The model is optimized for 4 steps. If you increase this value, ensure it is a multiple of 4 (e.g., 8, 12, 16) to prevent errors.

An example workflow is provided in the example folder. You can drag and drop the .json file from there into ComfyUI to get started quickly.

This project is simply a ComfyUI wrapper. All credit for the underlying technology and model architecture goes to the original authors of PersonaLive and the projects they built upon (Moore-AnimateAnyone, X-NeMo, StreamDiffusion, RAIN, LivePortrait).