Alibaba's Wan-Animate-2 brings any character to life

Alibaba has open-sourced Wan-Animate-2, an end-to-end character animation model that brings static images to life using driving videos. Feed it a single still photo and a template video, and it transfers full-body choreography, facial micro-expressions, and delicate finger gestures onto the target subject.

The project is available under the Apache 2.0 license, complete with inference scripts, 14B base checkpoints, and distilled weights for faster sampling.

What it does

Most existing image-to-video character animation tools rely on intermediate representations like skeleton pose maps or dense facial meshes. Those pipelines frequently lose identity details and result in jittery transitions. Wan-Animate-2 sidesteps this by rebuilding the Diffusion Transformer architecture to ingest the reference photo and driving video directly, ensuring much tighter identity preservation.

It handles nuanced movements exceptionally well, extending beyond broad torso motion down to finger gestures and lip sync. The subject does not have to be human either; stylized drawings, teddy bears, and cartoon animals with unconventional anatomy animate just as cleanly.

The framework also handles multi-person scenarios out of the box. You can map a two-person driving clip onto a two-character illustration, or use a solo reference video to animate multiple figures simultaneously. Furthermore, Wan-Animate-2 decouples camera angles via text prompts, allowing creators to shift camera perspectives independently from the original driving footage.

Specs and getting started

Alibaba released weights for the 14B base model as well as a distilled variant that cuts generation down to 10 steps without classifier-free guidance. A lightweight edition named Wan-Animate-2-Lite targets real-time streaming, pushing inference latency below one second for interactive setups.

Running the full model locally requires hefty compute. The official configuration is tuned for 8x A800 GPUs for 720p output, though 480p can run on 2x A800 setups. The weights take up around 33GB of storage, with INT8 versions cutting that footprint roughly in half.

Ecosystem support is solid right from launch, featuring integrations with Hugging Face Diffusers, DiffSynth-Studio, and ComfyUI. Those who prefer not to manage a local GPU stack can test out the online demo hosted on ModelScope Studio.

To run it yourself, set up Python 3.11 and PyTorch 2.7, install flash-attn, and clone the GitHub repository. Inference relies on a quick LLM pass to generate structured appearance and background captions before rendering the final animated clip.

Why it matters

Wan-Animate-2 represents a major leap in end-to-end motion transfer, putting high-fidelity digital performance capture into the hands of anyone with a standard camera clip. Indie animators, game developers, and VTubers can bypass expensive motion-capture rigs and turn static character artwork into expressive video assets.

Eliminating intermediate keypoint extractors makes character rendering look remarkably natural, matching the quality of top proprietary alternatives. The main hurdle remains local compute: a 14B model requires serious enterprise hardware, meaning most hobbyists will rely on ComfyUI INT8 quantization, cloud endpoints, or lightweight builds until consumer-grade optimizations catch up.