Overview
SCAIL-2 is an open-weight model for controlled character animation. It transfers motion from a driving video to a reference character and generates video, using direct visual conditioning rather than requiring an intermediate pose skeleton. The project is associated with Tsinghua University and Z.ai. (huggingface.co)
Capabilities and uses
- Animates a reference character, replaces characters, and supports multi-character animation and replacement.
- Supports cross-identity replacement, animal-driven motion and zero-shot use of certain mesh-rendering controls. Multi-reference inference is also available.
- Potential uses include character animation and video editing. Text prompts, reference images, driving videos and masks provide control. (github.com)
Limitations and access
- Correct masks are important: the repository warns that incorrect masks can cause animation mode to behave like replacement mode. Detailed prompts generally work better than short or empty ones.
- End-to-end driving supports 512p and 704p. For other resolutions, both height and width must be divisible by 32; the model card says pose-driven and replacement modes perform better at 704p.
- The weights and inference code are available for self-hosting, with ComfyUI integration. The GitHub repository states that its code is Apache-2.0 licensed, while the Hugging Face model card labels the model MIT; review the applicable terms separately. (github.com)
Release date
The official repository records the model and inference code as open-sourced on June 9, 2026. (github.com)
Pricing
As checked on November 16, 2026, official SCAIL-2 API pricing, including currency, billing unit and any separate input or output charges, could not be verified. The international pricing page does not list SCAIL-2; the mainland China pricing page could not be read without JavaScript. This does not establish that an API is free. Self-hosting can incur infrastructure costs. (docs.z.ai)