Wan Animate 2

Wan-Animate-2 (also listed as Animate Anyone 2 on the project page) is a character image animation framework from Alibaba's HumanAIGC group. It takes a reference character image and a driving video and generates a video of the character performing the motion. It uses a dual-branch Diffusion Transformer that consumes the driving video directly, without intermediate motion extractors, and introduces Time-Align RoPE and Sparse-Ref Attention to fuse appearance and motion cues.

Described capabilities include replication of body motion and facial expressions, multi-character animation (single-to-multiple and multiple-to-multiple driving), and text-driven viewpoint control that decouples the output camera angle from the driving video. Creator coverage also mentions handling of non-human characters, ComfyUI support and an int8 version.

Wan-Animate-2-Lite is a distilled variant trained with teacher forcing and Self-Forcing distillation, aimed at real-time streaming re-enactment from live camera feeds, for uses such as digital avatars and live-streaming hosts. The project page says the Wan-Animate-2-Base weights will be released; the provided material does not confirm download links, license or pricing, so those details are unknown.