SCAIL 2

SCAIL-2 is a character animation framework that transfers motion from a driving video to a reference character. It follows an end-to-end paradigm: latent visual information is concatenated directly into the input sequence of a latent video diffusion model, which removes the pose-skeleton and masked-background intermediates used by earlier methods. It is described as the successor to SCAIL (arXiv 2512.05905).

The project introduces MotionPair-60K, a dataset of 60K synthetic motion pairs generated with SCAIL-Preview, Wan-Animate and MoCha. It also introduces In-Context Mask Conditioning and Mode-Specific RoPE, which let a single model handle single-character animation, multi-character animation and character replacement. A post-training Bias-Aware DPO step targets synthetic-data errors in fine details such as hands and fingers.

The project page reports zero-shot generalisation to driving sources outside the training distribution, including animal-driven animation, cross-embodiment transfer and object motion transfer. It also reports emergent camera-following under large camera motion. A YouTube reviewer described it as open source, with weights totalling about 81 GB, and compared it favourably to Kling 3 and Wan Animate. The provided pages did not include a code repository, weights link or license, so these need to be checked. Vendor attribution to Z.ai comes from the listing and is not confirmed on the project page.