Alibaba Launches Wan-Animate-2 for Skeleton-Free Animation

Alibaba's Tongyi Lab has launched Wan-Animate-2, a fully open-source character animation framework that eliminates the need for intermediate pose skeletons or body meshes. Instead of warping a reference image to match an extracted skeleton—a process prone to losing fine details like hand movements and facial expressions—the new system feeds raw video latents directly into a redesigned Diffusion Transformer. This end-to-end approach preserves the original subject's identity and motion fidelity without the typical drift caused by structural differences between bodies.

The release includes a Lite variant designed for real-time streaming, which achieves 24 frames per second at a 400x720 resolution when running on four Nvidia H100 GPUs. In a blind user study, Wan-Animate-2 outperformed the proprietary platform Dreamina and matched the performance of Kling-MotionControl, despite both competitors relying on larger, closed-source models. Additionally, the framework supports multi-character animation, allowing developers to animate multiple distinct subjects within a single scene while maintaining their individual identities and movements.

For practitioners, the model introduces unprecedented creative control through a Viewpoint LoRA trained on approximately 50,000 Unreal Engine renders. This feature allows users to adjust the camera angle of the generated video using simple text prompts, decoupling the camera perspective from the original driving video. Because the model is released under the Apache 2.0 license, developers can access the weights on HuggingFace immediately. The system also integrates directly with popular ecosystem tools, including diffusers, ComfyUI, and DiffSynth-Studio, making it easy to adopt into existing production pipelines.