Wan2.2-Animate-2-14B

Wan2.2-Animate-2-14B is the second generation of Alibaba's character-animation model: weights and inference code went out on 7 August 2026 under Apache 2.0, with the paper posted a day later. The job it does is the same as before - hand it a still character, a drawing or a render or a photograph, plus a driving video of a person performing, and it makes the character reproduce that performance. What changed is the route the performance takes. The first generation, and almost every character animator before it, read the driving video through an intermediate motion representation: extract a pose skeleton, then hand the skeleton to the generator. The team's own paper names the two failure modes of that design without softening them - extraction errors propagate into the output, and the character's identity drifts as the clip runs. Animate-2 removes the middleman entirely. The driving video is consumed directly inside a redesigned diffusion transformer, which is why the paper calls the system end-to-end rather than merely improved. The second addition is easier to picture. Text-driven viewpoint control lets the prompt set the output camera independently of the camera that filmed the performance, so the same driving clip can be rendered from a different angle. Pose-based systems can rarely do this, for a structural reason: a skeleton extracted from footage carries the viewpoint of that footage baked into it, and nothing downstream can take it back out. The repository ships two checkpoints of the same size, 32.8 GB each in bfloat16 - roughly sixteen billion parameters by file size, despite the 14B in the name. The base build runs 40 denoising steps with classifier-free guidance; the distilled build reaches a picture in 10 steps with guidance switched off and a plain Euler solver. That is a quarter of the sampling work from an identically sized file. Alongside them sit the shared parts of the Wan stack: an 11.4 GB umt5-xxl text encoder, a 4.8 GB CLIP vision tower and the video autoencoder. Two things are worth knowing before downloading. First, the paper's headline claim is a real-time variant, Wan-Animate-2-Lite, trained in three stages - teacher-forcing pretraining with an error buffer, then Self-Forcing distillation with chunk-wise backpropagation - to bring latency down to streaming thresholds for live avatars. Lite is described in detail and is not in the repository, nor on the to-do list of things still to come; what has been released is the offline model. Second, the card links to a LICENSE file that the repository does not contain, the same gap that runs through the whole Wan2.2 generation, and the citation block still carries an unfilled placeholder where the arXiv number should be. Adoption so far runs through the conversions rather than the original. The base repository registers no downloads at all in its first thirty days against 185 bookmarks, while the Diffusers builds dated 6 August 2026 carry the traffic - and the distilled one is pulled roughly three times more often than the full-quality one, which says something about what people want from character animation. A repository that ships raw checkpoint files rather than a standard configuration often reads as zero on that counter, so treat the figure as a sign of how the model is being fetched rather than of how little. The first-generation Animate, meanwhile, is still being pulled an order of magnitude more than either.

▸News

Alibaba's new animation model was built for live video — and the live-video version is the one part it did not release

9/3/2026 · Hugging Face — karta modelu Wan2.2-Animate-2-14B; arXiv 2608.06009

▸Videos

No videos yet.

▸Reviews