SCAIL 2's feature set centers on removing the skeleton bottleneck from character animation. Every other feature follows from that design choice. Animals work because there is no human skeleton to violate. Multi-character works because there are no overlapping skeletons to misinterpret. Replacement works because the model reads visual identity directly, not through a skeleton intermediary.

End-to-End (No Skeleton)

SCAIL 2 concatenates driving video latents directly to the generation sequence. The model reads movement, appearance, environment, and fine visual details from the raw video. No skeleton estimation, no pose extraction, no intermediate representation at any stage. This is the core architectural difference from every other character animation model. Information that gets lost in a skeleton bottleneck (subtle head tilts, fabric motion, weight shifting, finger positioning) is preserved.

Four Modes in One Model

Single character animation, cross-identity replacement, multi-character scenes, and zero-shot animal/creature driving all run from the same model weights. No mode switching, no separate checkpoints, no different pipelines. The in-context mask conditioning and mode-specific RoPE handle the routing internally. Upload your inputs, and the model figures out what you are asking for.

Cross-Identity Replacement

Replace a character in a driving video with a completely different reference identity. The motion, timing, and scene composition transfer while the appearance changes. This is the Runway-style character replacement workflow, but open-source under Apache 2.0 and running through ComfyUI. Use it for talent swaps, brand character insertion, or creating localized variants of the same ad with different performers.

Multi-Character Group Motion

Animate multiple characters in the same scene. The model handles interaction, occlusion (one character passing behind another), and group motion coordination. Conventional models fail here because overlapping skeleton maps from multiple characters create ambiguous keypoint assignments. SCAIL 2 does not use skeletons, so there is nothing to overlap.

Animal and Creature Zero-Shot

Drive animal characters, pets, creatures, and fantasy figures from human or animal driving video. This works zero-shot because the model never depended on human skeleton anatomy in the first place. A cat, a robot, a slime creature, and a four-legged dragon all process through the same pipeline that handles human characters. No separate animal model needed.

Bias-Aware DPO LoRA

SCAIL 2 includes a DPO (Direct Preference Optimization) LoRA that corrects for bias patterns inherited from the pose-driven teacher models used during training. The DPO LoRA improves fine detail in hands, faces, and fabric. It is released separately on HuggingFace and can be enabled in both the standalone inference code and ComfyUI implementations.

MotionPair-60K Training Recipe

The model was trained on 60K synthesized motion pairs built from SCAIL-Preview, Wan-Animate, and MoCha. A "reverse driving" technique teaches the model to infer motion beyond what its teacher models could produce. This is why SCAIL 2 shows emergent capabilities (animal driving, zero-shot mesh rendering support) that none of its training sources had individually.