MagicAnimate

What MagicAnimate is

MagicAnimate is a free, open-source diffusion framework that animates a human reference image using a target DensePose motion sequence. An appearance encoder helps preserve the reference identity, while a video-fusion approach creates smoother transitions for longer animation inference. The project demonstrates cross-identity, unseen-domain, multi-person, and text-to-image-generated reference animation. It supports image-to-animation, motion transfer, and reference consistency. Users can try an online Gradio demo hosted on Hugging Face or launch a local Gradio demo from the repository on one or multiple GPUs. Local setup requires Python 3.8 or later, CUDA 11.3 or later, and ffmpeg. The documented model setup also requires Stable Diffusion v1.5, an MSE-finetuned VAE, and MagicAnimate checkpoints. The project is available for self-hosted and web use under the BSD-3-Clause license. Its authors are affiliated with Show Lab at the National University of Singapore and ByteDance.