Alibaba's Wan-Animate-2: Real-Time End-to-End Character Animation Framework

aigclink · x · 2026-08-10

Tongyi Lab at Alibaba introduced Wan-Animate-2, a new end-to-end character image animation framework. It directly consumes driving videos within a redesigned Diffusion Transformer, completely eliminating intermediate motion extractors (like skeletons or keypoints) to enhance motion fidelity and identity preservation.

The framework also features text-driven viewpoint control, decoupling the output camera perspective from the driving video. This overcomes previous limitations where generated videos were strictly bound to the shooting angle of the input.

To support real-time applications like digital avatars and live-streaming, the team released Wan-Animate-2-Lite. Using a three-stage training paradigm (teacher forcing with error buffer, followed by Self-Forcing distillation with chunk-wise backpropagation), it reduces inference latency to real-time levels, enabling streaming character animation.

Related event: Alibaba Open-Sources Wan-Animate-2 Character Animation Framework(4 posts)→

Original post →

More from Multimodal

Multimodal channel →