Alibaba’s new Apache 2.0-licensed Wan-Animate-2 framework delivers real-time, pose-free character animation that rivals top commercial AI tools.
Alibaba’s Tongyi Lab (formerly Tongyi Wanxiang) released Wan-Animate-2 on 7 August 2026, an end-to-end AI character animation framework whose model weights, inference scripts, and technical paper were made publicly available under the permissive Apache 2.0 licence on GitHub and Hugging Face.
To achieve real-time streaming at 24 frames per second (at 400×720 resolution), Alibaba introduced a specialised lightweight variant called Wan-Animate-2-Lite. Unlike traditional tools that rely on intermediate skeletal pose extraction or motion encoding, Wan-Animate-2 feeds raw driving video latents directly into a redesigned Diffusion Transformer (DiT) architecture using a dual-branch system with temporal RoPE alignment. This eliminates extraction bottlenecks, preserves fine-grained dynamics like hand movements and micro-expressions, and prevents identity drift.
Beyond pose-free animation, the system features text-driven viewpoint control via a specialised Viewpoint LoRA, allowing creators to decouple the camera angle from the driving video and render identical performances from different perspectives or with custom text-prompted backgrounds. The framework natively supports multi-character driving, non-standard aspect ratios, and diverse body types without needing retraining.
To lower hardware barriers for local deployment, the release includes quantised options (INT8 and BF16), distilled speed variants, and native integration nodes for creator workflows like ComfyUI and DiffSynth-Studio. Blind user studies show Wan-Animate-2 matching or surpassing closed-source commercial platforms like ByteDance’s Dreamina and Kuaishou’s KLING MotionControl, expanding on Alibaba’s broader push into enterprise AI video generation following its Wan 3.0 release earlier in August 2026.
