ZAI drops SCAIL-2 for extreme video motion transfer
ZAI just open-sourced SCAIL-2, an end-to-end character animation model that transfers complex motion without relying on ambiguous skeleton maps. It handles intense multi-character combat and can even map human dancing onto flamingos or stylized cartoons. The catch: the full weights clock in at 81GB, so consumer GPUs will likely need to wait for quantized builds.
SCAIL-2 is an open-source character animation model from ZAI that transfers motion from one video to another completely end-to-end. It skips the traditional intermediate step of extracting stick-figure skeletons, moving motion directly across frames.
That means no clunky rig retargeting or broken hands caused by faulty pose detectors. You feed in a driving clip and an image or character prompt, and the model handles the mapping at 512p or 704p.
What it can do
Its biggest party trick is handling violent, chaotic multi-character interaction. In test clips featuring intense two-person hand-to-hand combat, SCAIL-2 tracks both fighters simultaneously, faithfully reproducing heavy grappling and fast limb exchanges without blending the characters together.
It also handles non-human bodies surprisingly well. In one test, the system mapped a dancing human onto a flamingo, naturally translating human limbs into avian anatomy while preserving the rhythm and trajectory.
Stylized anime designs and wildly proportioned creatures work just as smoothly. The model figures out where the limbs belong even if the character is unnaturally round or distorted. It also faithfully preserves complex driving camera sweeps, an area where leading closed-source video tools often drift.
How it works
Older animation tools lean heavily on openpose maps or inpainting masks. Those intermediate representations break down whenever limbs cross, scenes get dark, or the subject simply is not human.
SCAIL-2 abandons intermediate poses altogether. ZAI built an in-context conditioning architecture using a Unified Motion Transfer Interface with specialized masking channels and custom RoPE position encoding. They trained it on 60,000 synthetic motion pairs and added Bias-Aware DPO to stamp out hallucinations and refine textural details.
The system uses color-coded masks to signal visible backgrounds and cross-character correspondences. This setup even enables zero-shot multi-reference generation to fill in occluded angles like character backs.
Requirements and setup
The code, inference scripts, ComfyUI integration, and training pipeline are all live on GitHub. The Hugging Face repo packages everything together, including Wan VAE and T5 text encoders.
The catch is raw size. The complete model weights weigh in at roughly 81GB. Running inference locally will crush typical consumer GPUs, so unless you have access to enterprise hardware, you will want to wait for community quantizations or GGUF builds before downloading.
Why it matters
For animators, indie game developers, and VFX creators, motion retargeting has always been tedious, especially across disparate body types. SCAIL-2 proves that end-to-end generative models can replace brittle skeleton tracking without losing precision.
Matching the visual stability of top-tier closed models while fully open-sourcing the training pipeline puts serious pressure on proprietary video platforms. Once lightweight quants arrive on ComfyUI, this could easily become the default motion engine for open AI filmmaking.
Sources (1)
- github.com(opens in a new tab)Originalgithub.com