MagicPose is an identity-aware diffusion model for realistic human pose and facial expression retargeting. It takes reference poses and transfers them to different character identities while maintaining natural motion and expressions. The model is published at ICML 2024 and includes implementation for pose-driven video generation.