UniAnimate

UniAnimate is a research framework for consistent human image animation using unified video diffusion models. It maps reference identity images, pose guidance, and noise into a common feature space, eliminating the need for extra reference models. The framework supports both random noised input and first frame conditioned input to enable long-term video generation (up to one minute). It employs a state space model (Mamba) for efficient temporal modeling instead of computation-heavy Transformers. The system uses CLIP and VAE encoders for reference features, a pose encoder for driven poses, and a unified diffusion model for denoising. Experimental results show superior synthesis over existing methods in both quantitative and qualitative evaluations.