Wan Animate 2 Character Animation

What is Wan Animate 2?

Wan Animate 2 is an end-to-end character-animation framework that reads a driving video directly instead of first reducing the performance to a pose skeleton. Its dual-branch architecture is designed to preserve appearance while carrying motion, expression and timing into the generated character.

• Official inputs are a reference character image, a driving video and a text prompt.
• The research supports human, cartoon, animal and other stylized subjects with different body proportions.
• The official release includes Base and Distillation checkpoints plus inference code.

Transfer a performance without losing the character

A single driving clip supplies the body movement, facial performance and rhythm while the reference image supplies identity and visual design. Clean framing and readable movement make the source easier to follow.

Animate one or several characters

The official project demonstrates single-to-multiple and multiple-to-multiple motion transfer. It can map coordinated performances into scenes with more than one subject while keeping their visual roles distinct.

Change the camera viewpoint with text

Wan Animate 2 decouples character motion from the source camera. A prompt can request a front, side, overhead or other viewpoint while the generated performance follows the driving video.

Wan-Animate-2-Lite targets real-time streaming

The research also introduces a Lite variant for streaming character animation. This is an official research capability, not a promise that every hosted service or local GPU will reach the same frame rate.

Open source and local deployment

The official Apache-2.0 repository publishes Base and Distillation weights under the model identifier Wan2.2-Animate-2-14B. The documented 720p configuration is tuned for eight A800 GPUs; the team also reports testing 480p on two A800 GPUs, so local deployment is aimed at high-memory multi-GPU systems.

• The reference environment uses Python 3.11, PyTorch 2.7.0 and FlashAttention.
• Base inference uses 40 sampling steps; the distilled checkpoint uses 10.
• The official repository does not publish a GGUF package or a consumer-GPU recipe. Treat third-party GGUF files as unofficial conversions and verify their license and provenance separately.

Why the download is called Wan2.2-Animate-2-14B

The 2026 project is branded Wan Animate 2, while the released model directory retains the Wan2.2-Animate-2-14B name. It is a newer direct-driving-video system, not merely the older Wan2.2 Animate page with a renamed heading.

Use the lower-cost online workflow

The embedded WAN3 tool below is a compatible reference-to-video workflow powered by Seedance Mini; it does not run the official Wan Animate 2 open weights. Add one character image and one motion video, choose a 4–15 second 720p output and review the credit total before submitting.

• Mini costs 17.5 credits per output second, rounded to the nearest five credits; four seconds costs 70 credits in WAN3.
• Use a clear full-character image and a driving clip with one readable performance and limited cuts.
• The server locks this page to the Mini reference workflow even if browser request fields are changed.