Champ

Champ animates a person in a single reference image by extracting SMPL parametric body sequences from a driving video and rendering depth maps, surface normals, semantic segmentation, and DWPose skeletons as conditioning signals for a latent diffusion generator. Because the guidance comes from a unified 3D body model rather than 2D keypoints alone, body shape and motion stay consistent even under large pose changes, which was the core contribution of the ECCV 2024 paper. The project comes from Fudan University's generative vision lab, the same group behind the Hallo talking-head models, and ships pretrained weights on Hugging Face along with full two-stage training code and sample training data. Running inference requires Python 3.10, CUDA 12.1, and roughly 20 GB of VRAM for a 250-frame sequence on an A100 or RTX 3090, with a frame-range option to trim memory use on smaller cards. Community wrappers exist for ComfyUI and Blender. Code and weights are MIT licensed, so commercial use is permitted, and the repository holds about 4,300 GitHub stars.