MagicAnimate

Give MagicAnimate one photo of a person plus a DensePose motion sequence, and it renders a video of that person performing the motion while keeping face, clothing, and background stable from frame to frame. Researchers from the National University of Singapore and ByteDance built it on Stable Diffusion 1.5, adding an appearance encoder that carries detail over from the reference image, a DensePose ControlNet that steers body position, and temporal attention layers that smooth movement across frames. The repository offers inference scripts for one or several GPUs and a Gradio demo you can run locally, with checkpoints fetched from Hugging Face next to the SD 1.5 base model and a fine-tuned VAE. Plan on Python 3.8 or newer, CUDA 11.3 or newer, ffmpeg, and a capable NVIDIA GPU. Motion must be supplied as a DensePose video, which the repository does not create for you. The code is BSD 3-Clause, but the SD 1.5 base weights carry the CreativeML OpenRAIL-M license, so review both before commercial use. Apart from small demo fixes in 2025 the code has not changed since December 2023, so it suits research and hobby use more than production.