Turn one character image and a driving video into expressive animation with accurate motion, stable identity, and flexible camera direction.
UPLOAD
JPG, JPEG, PNG
Max 10MB, Min 300×300px
Aspect ratio: 1:2.5 ~ 2.5:1
UPLOAD
MP4, MOV
Max 100MB, Min 3s
Resolution: 340px ~ 3850px
Max 120s
A direct video-conditioned framework built for detailed performance transfer across very different character types.
Earlier animation systems often reduce a driving performance to poses or compressed motion features. That can lose small expressions, hand details, and non-rigid movement.
Wan Animate 2 sends the driving video through a dedicated branch of the Diffusion Transformer. Frame-aligned attention connects that motion to the target character while keeping the two streams efficient.
The result is designed to preserve identity and motion across people, cartoons, animals, robots, close-ups, and full-body performances.
Wan-Animate-2 achieves superior visual fidelity with enhanced dynamic details, enabling the replication of intricate motions, nuanced facial expressions, and physically plausible character-scene interactions. It exhibits exceptional robustness and stability across diverse input combinations of varied aspect ratios and character morphologies.
Wan-Animate-2 enables controllable multi-person video synthesis, supporting versatile animation paradigms such as single-to-multiple and multiple-to-multiple motion driving.
Wan-Animate-2 enables view-controllable synthesis, allowing for multi-angle observations of the same character animation via explicit camera pose conditioning. Since motion is not tied to rigid spatial anchors, the viewpoint can be effectively decoupled from the reference information, ensuring robust multi-view content consistency.
Wan-Animate-2-Lite enables real-time streaming character re-enactment directly from live camera feeds, ensuring seamless synchronization of both motions and facial expressions with minimal latency.
Start with appearance, add performance, then describe how the camera should see the result.
Upload one clear image that defines the subject, style, clothing, and visual identity.
Use a driving clip with the movement, expression, and timing you want to transfer.
Describe the desired camera angle, generate, and review the result inside the creation workspace.
Use performance transfer where consistent characters and readable movement matter most.
Build expressive hosts for explainers, product education, and virtual events.
Turn mascots, illustrations, and original characters into repeatable content.
Test performance, framing, and character direction before full production.
Explore low-latency characters for live streams and real-time environments.
Clear answers about inputs, motion fidelity, viewpoint control, and the real-time research variant.
Wan Animate 2 is an end-to-end character animation framework that uses a character image and a driving video to generate a new performance while preserving the target character identity.
The core workflow uses a reference character image and a driving video. The image defines the subject, clothing, appearance, and visual identity, while the driving clip supplies body movement, facial expression, hand motion, and timing for the generated performance.
It directly consumes the driving video inside a redesigned Diffusion Transformer instead of depending on an intermediate pose or motion extractor. This avoids information loss and extraction errors described in the research paper.
The published examples focus on fine-grained facial expressions, hand articulation, full-body movement, and non-rigid motion across humans, animals, robots, and stylized characters.
Yes. Its optional Viewpoint LoRA accepts text descriptions such as front, side, top, bottom, and eye-level views, allowing the output camera to differ from the driving video. This separates performance direction from viewpoint direction so one motion can be explored from several angles.
Wan Animate 2 Lite is the research team’s accelerated streaming variant. It generates video in temporal chunks for interactive digital humans, virtual environments, and live-streaming applications.
The paper reports 24 fps at 400 by 720 resolution on a four-GPU NVIDIA H100 research setup. Real-world speed depends heavily on the implementation and hardware, so this result should not be treated as a consumer-device guarantee.
Typical uses include animated brand characters, digital presenters, stylized dance clips, coordinated character scenes, virtual production tests, game and film previsualization, and interactive avatar experiences. It is especially useful when a project needs a recognizable character to follow a specific recorded performance.
Yes. The published results demonstrate controllable multi-person synthesis, including single-to-multiple and multiple-to-multiple motion-driving setups. This supports paired performances, coordinated groups, and scenes where several characters need to follow one or more driving sources.
The published examples cover realistic people, illustrations, animated characters, animals, and robots with varied proportions and appearances. Output quality still depends on how clearly the character is shown and how well the driving motion can be read, so highly occluded or ambiguous inputs may be more difficult.
Choose a clear character image that shows the subject’s identity, clothing, and body structure. For the driving video, use readable movement with the face, hands, and key body actions visible when those details matter. Reducing heavy occlusion and abrupt framing changes gives the model more consistent motion information to follow.
Wan Animate 2 uses viewpoint conditioning to separate camera pose from the reference performance and is designed to maintain the same character across multiple views. The research examples include front, left, right, top, and bottom viewpoints. This provides directional control, although it should not be treated as a guarantee of exact frame-by-frame geometry.