SCAIL-2: Tsinghua's Diffusion Framework Animates Characters End-to-End Without Skeleton Intermediates

Researchers at Tsinghua University and Z.ai have released SCAIL-2, a diffusion-based framework for controlled character animation that eliminates reliance on skeleton maps, character masks, and other intermediate representations. By concatenating driving video latents directly into the model sequence, SCAIL-2 transfers motion from a source to a reference character in a single end-to-end pass — handling single-character animation, cross-identity replacement, and multi-character interactions within one unified architecture.

To address the scarcity of end-to-end training data, the team synthesized MotionPair-60K, a 60,000-pair dataset spanning animation, replacement, and multi-character tasks. Three key contributions underpin the system: in-context mask conditioning that switches between environment and character-binding modes, mode-specific RoPE position encodings that route spatial-temporal attention per task, and Bias-Aware DPO — a post-training refinement step targeting synthetic-data artifacts concentrated in fine-grained regions such as hands and fingers. The authors report that SCAIL-2 outperforms current state-of-the-art pose-driven methods and also unlocks zero-shot generalization to animal-driven animation and mesh-based control.