Zhipu AI releases SCAIL-2 to industrialize controlled character video generation

SCAIL-2 addresses the primary bottleneck in AI video production: the lack of granular, predictable control required for professional film and television. By assetizing performance through reusable visual vectors rather than abstracted skeletons, Zhipu is attempting to shift the production pipeline from manual character rigging to intention-driven creation. For the broader industry, this moves Zhipu from a tool provider to an infrastructure layer offering a standardized production protocol for digital assets. Competitively, it positions Zhipu as a peer to North American giants like Runway by providing open-source access to high-end control mechanisms. Watch for Zhipu's integration into major game engines and traditional VFX software as a signal of industrial adoption.

Zhipu AI's SCAIL-2 model was released under the MIT license on June 9, 2026, with inference code and model weights published simultaneously on GitHub and Hugging Face. The model supports end-to-end character animation including character replacement and multi-character scenarios without intermediate pose representations, and within days of launch the team shipped ComfyUI integration on June 9, multi-reference inference on June 13, and multi-reference ComfyUI support on June 17. A relighting LoRA for replacement mode followed on July 15, 2026, extending the model's utility for compositing workflows where lighting consistency between replaced characters and background plates is critical.

The technical paper, authored by researchers at Zhipu and Tsinghua University's KEG Laboratory under Professor Tang Jie, describes a data pipeline that synthesized 60,000 motion pairs using off-the-shelf models including SCAIL-Preview, Wan-Animate, and MoCha. The team employed an agentic synthetic loop to curate diverse high-quality animation pairs and used reverse driving so the final model learns capabilities beyond its teacher models, yielding emergent abilities such as multi-character interaction and environment-aware motion transfer. A Bias-Aware DPO training stage addresses synthetic discrepancies in detailed regions, and the authors report that SCAIL-2 substantially outperforms existing state-of-the-art approaches across various character animation benchmarks.

The Open Weights reported that SCAIL-2's permissive MIT license allows broad commercial adoption and encourages developers to integrate it into new applications, positioning the model as a specialized tool for animators, game developers, and creative technologists rather than a general-purpose text-to-video system. Zhipu's strategy of combining open-source weights with ComfyUI workflow integration targets the same creator ecosystem that has driven adoption of other open video models, but with a focus on pose-driven control that gives directors granular command over character movement rather than relying on text prompts alone.