SCAIL-2: End-to-End Character Animation With Native ComfyUI Support

SCAIL-2 (HuggingFace | GitHub | Project Page) is Zhipu AI's open-source character animation model: it animates a reference character with a driving video end-to-end, without intermediate pose representations, and ships with official native ComfyUI support merged into ComfyUI core.

Overview

SCAIL-2 is built on the WAN2.1 14B architecture and removes the dependence on intermediate representations such as skeleton maps or inpainting masks that earlier character animation models rely on. The team synthesized 60K motion pairs from off-the-shelf models (SCAIL-Preview, Wan-Animate, MoCha) and trained through a Unified Motion Transfer Interface with dedicated masking channels and a RoPE design.

SCAIL-2 end-to-end character animation with cross-identity replacement and multi-character scenarios (source: project page)

The reverse driving training recipe lets the model learn capabilities beyond its teacher models, yielding emergent abilities:

Cross-identity character replacement — animate a different character with the same driving motion
Animal-driving scenarios — drive a character with animal motion sources
Zero-shot advanced control — support for intermediate controls like SAM3D-Body mesh rendering without retraining

ComfyUI Support (Native)

SCAIL-2 is integrated directly into ComfyUI core: the initial support PR (Comfy-Org/ComfyUI #14373) merged on the release day, and multi-reference support followed in PR #14509. No custom node pack is required.

Comfy-Org repackaged weights provide ready-to-use diffusion models in fp16, FP8, INT8, MXFP8, and NVFP4/MXFP8 mixed precision, plus the Bias-Aware DPO LoRA (fixes hand distortion, improves lip/eye sync) and the Relighting LoRA (blends replacement characters into target lighting).

SCAIL-2 network architecture: Unified Motion Transfer Interface with masking channels and RoPE (source: project page)

Demo Videos

SCAIL-2 teaser samples (video from project page)

Single-reference animation comparison (video from project page)

Workflow

The SCAIL-2 example workflow from Kijai's ComfyUI-WanVideoWrapper is a good starting point:

For multi-reference mode, the official test workflow from Comfy-Org's integration PR is also available:

Key Details

Item	Detail
Resolutions	End-to-end driving: 512p and 704p; pose-driven performs better at 704p
Constraints	H and W must both be divisible by 32 (e.g. 704x1280)
Bundled modules	Wan VAE and T5 integrated into the checkpoint
Extras	Bias-Aware DPO LoRA, Relighting LoRA, training code (2026-08-06)