Wan-Animate-2: End-to-End Character Animation With ComfyUI Support
Alibaba Tongyi Lab releases Wan-Animate-2, an end-to-end character animation model consuming driving videos directly, with text-driven viewpoint control and native ComfyUI nodes.
Alibaba Tongyi Lab released Wan-Animate-2 on August 7, 2026 — an end-to-end character animation framework that consumes a driving video directly inside a redesigned Diffusion Transformer. The Base and Distillation model weights, inference scripts, and an arXiv paper are all public under the Apache-2.0 license.
Wan-Animate-2 dual-branch DiT architecture: the driving video is consumed directly, with Time-Align RoPE and Sparse-Ref Attention fusing appearance and motion (source: project page)
Demo Videos
Single-person high-fidelity character animation
Multi-character animation: one-to-multiple and multiple-to-multiple motion driving
Real-time streaming character animation with Wan-Animate-2-Lite
What's New vs Wan-Animate
- No intermediate motion extractors. Wan-Animate-2 feeds the driving video directly into a redesigned Diffusion Transformer, eliminating skeleton/expression extraction steps that caused identity drift and motion errors.
- Text-driven viewpoint control. The output camera perspective is decoupled from the driving video, enabling multi-angle shots of the same animation with consistent identity.
- Multi-character animation. Supports single-to-multiple and multiple-to-multiple motion driving in one pass.
- Wan-Animate-2-Lite. An efficient variant that reaches real-time inference thresholds for streaming character animation, aimed at digital avatars and live-streaming hosts.
How It Works
Wan-Animate-2 uses a dual-branch DiT architecture that conditions generation on a reference image and a reference video at the same time:
- Time-Align RoPE aligns temporal positions between denoising video tokens and reference tokens.
- Sparse-Ref Attention selectively attends to informative reference features instead of attending to the full reference sequence every step.
The distillation variant runs 10 steps with no classifier-free guidance (Euler scheduler), while the Base model follows the standard 40-step sampling.
ComfyUI Support
Wan-Animate-2 is supported natively in ComfyUI with two new experimental nodes: WanAnimate2ToVideo (animates a character from a reference image using a pose/driving video, with pose_strength, pose_start_percent, pose_end_percent, and reference_image_strength controls) and WanAnimate2Cache (caches the pose branch's per-block activations, roughly halving generation time at the cost of ~12.5 GB system RAM at 480x832/81 frames).
The Comfy-Org repackaged weights provide bf16 and int8-convrot diffusion models, the lightx2v step-distillation LoRA, text encoders, VAE, and CLIP vision in ComfyUI-ready folders. The official workflow template:
Comments
Sign in with GitHub to join the discussion.