WeChatCV /Stand-In

About Stand-In

Stand-In adds a consistent face to Wan video generation from one photo. It trains only about 1% extra parameters, so the base model stays frozen. You get scripts for text-to-video, LoRA styles, face swapping and VACE pose control. Training code and dataset aren't out yet, and face swapping is experimental.

Best for: Technical artists who need one face to stay consistent across Wan-generated shots.

• Keeps a character's face the same across shots from one reference photo
• Trains about 1% extra parameters and leaves the base model frozen
• Loads community LoRA (small add-on style models) for stylized video
• Swaps faces in existing video with a denoising-strength setting
• Works with VACE (a control model) for pose or depth control plus identity
• Downloads base, face-recognition and Stand-In weights in one script

1. 1
   Clone the repo and make a Python 3.11 environment: the command below, the next command, conda activate Stand-In

   git clone https://github.com/WeChatCV/Stand-In.git

   Copy

   conda create -n Stand-In python=3.11 -y

   Copy
2. 2
   Install the requirements: the command below

   pip install -r requirements.txt

   Copy
3. 3
   Download the weights into checkpoints: python download_models.py (add --wan_version 2.2 for Wan2.2, --vace for VACE)
4. 4
   Generate a video: the command below

   python infer.py --prompt "..." --ip_image "test/input/lecun.jpg" --output "test/output/lecun.mp4"

   Copy
5. 5
   For a style, run python infer_with_lora.py with --lora_path and --lora_scale
6. 6
   For face swapping, run python infer_face_swap.py with --denoising_strength