Video generation models have advanced rapidly in quality and length over the past year, but character animation has remained stuck in the old "extract skeleton first, then drive" paradigm. Zhipu AI and Professor Liu Yongjin's group at Tsinghua jointly open-sourced SCAIL-2, cutting that paradigm open.
SCAIL-2's core breakthrough is the complete abandonment of explicit intermediate representations like 2D keypoints and SMPL Mesh, directly stitching together the latent-space features of the driving video and the reference character at the pixel level, letting the model understand motion through "visual intuition" rather than "symbolic translation." Combined with full-context pose injection in the DiT architecture and Pose-Shifted RoPE, the model handles scenarios where traditional approaches nearly fail — multi-person complex interactions, zero-shot animal driving generalization. SCAIL-2 supports 512p/704p dual resolution, ships under Apache 2.0, with weights on Hugging Face, ModelScope, and GitHub, and ComfyUI workflows ready out of the box.
The deeper engineering significance is the compute simplification that comes with end-to-end: traditional pipelines require multiple serial steps — skeleton extraction, pose re-projection, mask generation — and SCAIL-2 folds them all into a single Transformer, significantly reducing inference latency and VRAM. Zhipu's "AI synthesizes AI data" factory-style pipeline is letting character motion evolve from "stick figures" to reusable visual vectors, with direct commercial value for games, livestreaming, and film/TV digital-human production lines.
SCAIL-2 still has its limits: fine-grained control of hands and faces still depends on large-scale high-quality paired data. But the "industrial-grade precise control" path is much closer to actual production-tool needs than simply scaling parameters — and it's a key signal that video generation is industrializing in 2026.