Wan-Animate-2 — Alibaba Open-Sourced the Skeleton-Free Model That Makes Your Still Character Act, in Real Time (and Beats the Paid Ones)

Wan-Animate-2 turning a still character into a moving performance
Wan-Animate-2 drives any still character from a single reference video. Source: Wan-Animate

Alibaba just open-sourced the animation tool a lot of us have been faking with duct tape. Wan-Animate-2 takes one still character and one driving video, and makes the character perform the video — face, hands, body, the works. No skeleton rig. No mocap suit. And the lite build runs in real time on your own machine.

The Story

Tongyi Lab dropped Wan-Animate-2 on August 7, 2026, under the Apache 2.0 license. Weights, inference code, and a paper — all public, all commercial-friendly. It does two jobs from the same model.

  • Animation mode: feed it a static character image plus a driving video. The character copies every move and expression in that video.
  • Replacement mode: drop your character into an existing video, swapping out the person who was there — and it relights your character to match the scene.
Animation mode and replacement mode side by side
One model, two jobs: reenact a character (left) or replace one inside a real clip (right). Source: Wan-Animate paper

Here is the part that matters for quality. Older tools first pull a stick-figure skeleton out of the driving video, then animate from that skeleton. Every subtle thing — a finger curl, a raised eyebrow, a lip twitch — gets flattened into a few joint dots and lost. Wan-Animate-2 skips that step. It feeds the raw video straight into a diffusion transformer, so the fine motion survives. Faces and hands stay alive instead of going rubbery.

Wan-Animate-2 architecture diagram
The pipeline injects motion and implicit face features into a Wan diffusion backbone, with a relighting step for clean compositing. Source: Wan-Animate paper

Why You Should Care

Two numbers make this real. First, the Lite variant streams at 24 fps at 400×720 — real-time, on a single consumer GPU. Second, in blind user tests it matched or beat the closed platforms people actually pay for: ByteDance Dreamina and Kuaishou KLING MotionControl. Free and open, winning against paid and locked.

For 3D and character artists, this fills an ugly gap. You can generate a gorgeous character in seconds now — Tripo, Rodin, Meshy all do it. But making that character act still meant rigging, weight painting, and mocap cleanup. Wan-Animate-2 lets you act it out on camera and hand the performance to your character directly. Concept-to-motion in one pass.

The replacement mode is a VFX cheat code too. Shoot a scene with yourself as a stand-in, then swap in your hero character with matched lighting. No greenscreen, no keyframe roto.

Diverse Wan-Animate results across styles and motions
Cross-style transfer, complex motion, and character replacement from the same model. Source: Wan-Animate paper

Try It / Follow Them

It ships with native ComfyUI and DiffSynth-Studio nodes, plus INT8 and BF16 quantized builds. So you don’t need to touch a command line — pull the ComfyUI template, load your character image and a driving clip, and hit run.

IK3D Lab Take

The dead-statue problem — beautiful AI characters that can’t move — has been the quiet wall in this whole pipeline. Wan-Animate-2 knocks a big hole in it, and does it in the open. Skeleton-free driving is the right call: skeletons were always a lossy middleman. The honest caveat is that this outputs video, not a rigged 3D asset — you’re driving a 2D performance, not baking animation onto your mesh. But paired with the image-to-3D wave, it’s the fastest path yet from a character in your head to a character on screen, acting. We’re loading the ComfyUI nodes this week.

Sharing is caring!

Leave a Reply

Your email address will not be published. Required fields are marked *