champ

Why we included this project

Champ takes a single photo of a person and turns it into a short video clip, with the movement supplied by a reference motion sequence rather than a text prompt. It conditions a latent diffusion model on 3D parametric guidance from the SMPL body model, along with depth, pose, normal, and semantic maps, which keeps the animated person recognizable and proportionally correct as they move. That level of control is what makes it attractive for avatar and dance-video work: you provide a reference image and a target motion sequence, and the model re-renders that specific person following the motion. The repo carries pretrained weights, an inference script, training code, and docs for producing your own SMPL motion data, so teams can run it locally and adapt the pipeline instead of treating it as a black box. For character animation or digital human projects, it is a solid research-grade baseline from an ECCV 2024 paper.