DreamActor-M1: Holistic Human Image Animation

DreamActor-M1 is a novel framework for generating realistic human animations from a single image by learning from videos. This system employs a diffusion transformer (DiT) architecture and utilizes a hybrid guidance approach combining facial features, head positioning, and body skeletons for precise and expressive control. Its design enables adaptability across different scales, from close-up portraits to full-body movements, while maintaining long-term visual consistency and likeness to the reference image. Through innovations like progressive training and integrated appearance cues, DreamActor-M1 aims to overcome limitations in existing animation methods, offering enhanced controllability and robustness.