What is scail?

SCAIL is a research model for animating characters in video. Given a reference image of a character and a description of a pose or motion, it generates a video of that character performing the motion, while keeping the character's identity and appearance consistent throughout. The project is the official code release for a paper accepted at CVPR 2026, a major computer vision research conference. The core technical idea behind SCAIL is what the authors call three dimensional consistent pose representation. Earlier character animation methods often struggled with generalizing to unfamiliar characters and produced awkward or physically implausible motion during complex actions like turning or flipping. Instead of just telling the model what pose to follow, SCAIL shows the model fuller context about depth and motion, which the authors found helps even smaller models learn to generate more believable movement. The technique is described as in context learning, meaning the model learns from examples shown directly in its input rather than through additional fine tuning. SCAIL is built as a fourteen billion parameter diffusion transformer, a type of AI model architecture, and the preview checkpoint is available for download from Hugging Face and ModelScope. Community members have reported it generalizing surprisingly well beyond its training data, including to anime style characters, hand drawn artwork, and even four legged animals, despite very little of that kind of data being in the training set. There is also a ComfyUI integration for people who prefer a visual workflow over writing code. Because this is a research preview rather than a finished product, running it requires downloading multi gigabyte model checkpoints and a GPU capable of handling a fourteen billion parameter model. The project provides inference code only, training code for reproducing the full pipeline was still being prepared for release at the time of writing. The full README is longer than what was shown.