Character animation model from Z.ai that runs on a rival's backbone: the weights ship with Alibaba's Wan components and the default code branch is called wan-scail2.
SCAIL-2 animates a still character image with a driving video. It also replaces a character inside existing footage and handles scenes with several characters at once. What sets it apart from earlier work is that it drops the intermediate step: competing systems first convert a performance into a skeleton map or an inpainting mask, then feed that to the generator. SCAIL-2 conditions on the driving video directly, which is why it can be driven by an animal or by a mesh render it was never trained on.
The unusual part is where the model comes from. Z.ai publishes flagship language models of its own, but this one stands on a competitor's foundation: the download bundles Alibaba's Wan 2.1 VAE and Google's umT5-XXL text encoder, the inference command loads a checkpoint named SCAIL-14B, and the default branch of the code repository is called wan-scail2. The training data was synthesised the same way, with 60,000 motion pairs generated by three off-the-shelf models, two of them from other labs: Alibaba's Wan-Animate and Orange's MoCha. The company describes this as reverse driving, and reports that the result learned behaviour its teachers never had.
The licence is declared twice and differently. The code repository is Apache 2.0. The weights repository carries an MIT label in its metadata and no licence file at all, even though the archive it distributes contains third-party components released under Apache 2.0. Anyone redistributing the checkpoint should read both notices rather than the badge.
#video generation#character animation#open weights#image to video