depth_anything_v2

Depth Anything V2

The VACE WAN model requires pretrained weights for Depth Anything V2 to extract depth information from video frames.

Download Instructions

Download the pretrained weights from the official repository and place them in ${CKPT_ROOT_DIR}/annotators/:

mkdir -p ${CKPT_ROOT_DIR}/annotators
wget -O ${CKPT_ROOT_DIR}/annotators/depth_anything_v2_vitl.pth \
    "https://huggingface.co/depth-anything/Depth-Anything-V2-Large/resolve/main/depth_anything_v2_vitl.pth?download=true"

Configuration

The experiment configs use CKPT_ROOT_DIR by default:

config.model.net.depth_model_path = f"{CKPT_ROOT_DIR}/annotators/depth_anything_v2_vitl.pth"

You can override CKPT_ROOT_DIR via environment variable or specify a custom path directly.

Notes

• The depth model is used by VACE (Visibility-Aware Context Encoding) to enhance video generation
• The model will be loaded automatically when running VACE WAN training or inference
• Ensure the weights file has proper read permissions

The module is imported by networks/VaceWan/modules/vace_depth_annotator.py and is essential for VACE WAN functionality.

License

See licenses/depth_anything_v2/LICENSE for details.