SCAIL video model

SCAIL means Studio-Grade Character Animation via In-Context Learning. It’s a character animation model that takes motion from a video or pose sequence and applies it to a single reference image. The result is an animated video where the character moves in a realistic way and stays visually stable over time.

Older tools lean mostly on 2D keypoints. SCAIL takes a different route and uses a 3D-consistent pose setup that captures deeper motion details. This helps create smoother motion, cleaner body rotations, and stronger results during complex actions like sharp pose changes, scenes with more than one character, or motion borrowed from a different identity.

Key things it can do.

Pose-guided animation.You animate one image using motion from another video or pose track.

3D-based motion handling.The 3D pose setup avoids common confusion that comes with 2D keypoints.

Smooth motion over time.Looking at the full sequence helps motion stay steady across many frames.

Diffusion transformer core.This setup supports higher quality motion output.

Open source access.The code is public under Apache 2.0 with training and testing tools included.

Some early feedback on Twitter says the model still struggles with consistency. So it works, but it’s not perfect yet.