A video-generation technique that extracts motion from one clip and applies it to a different subject — a still image or a generated character, no mocap gear.
Last verified · 2026-11-03 · by Moe Ameen
AI motion transfer is a video-generation technique that moves the motion out of one clip and onto a different subject. You supply a reference — commonly called a driving video or a conditioning video — of something moving, and a target such as a still photo, an illustration, or a generated character. The system outputs new video in which the target performs the reference's motion: the dance you filmed comes back performed by a cartoon character, a mascot, or a person who was never in the room. It is the shared technique underneath a wave of 2026 features, from Runway's performance-capture apps to Kling's Motion Control, ByteDance's DreamActor, and the open-source Wan Animate.
Most systems run two stages. First, a pose estimator tracks anatomical keypoints — joints, limbs, body contour, and in newer systems hands and face — across every frame of the driving clip, producing a pose or skeleton sequence that abstracts the movement away from whatever the reference subject looked like. Second, a generative model, almost always a diffusion model guided by a ControlNet-style conditioning signal, renders the target following that pose sequence, using an embedding of the target's appearance to keep it recognizable. Temporal-attention modules fight to keep the character coherent between frames, because holding a character's identity stable while the body does something new is the field's central challenge. Some 2026 models, notably ByteDance's DreamActor, are pose-free and skip the explicit skeleton.
Two adjacent terms get confused with it. Motion transfer is not motion capture: mocap uses sensor suits and markers and outputs riggable animation data, while transfer reads movement from ordinary video and outputs a finished clip. And it is not [video-to-video](/glossary/photo-to-video-ai) restyling: restyling keeps the original subject and motion and changes the look, while transfer keeps the motion and changes the subject. The clean test is to look at who is in the output — a different character performing the input's movement is motion transfer; the same subject in a new style is video-to-video.
The building blocks arrived in open source before the polished products did. Pose-conditioned image generation via ControlNet and temporal coherence via the AnimateDiff lineage gave the research community a way to drive a generated subject with an extracted skeleton, and early motion-transfer workflows were stitched together from those pieces. The capability existed, but it took real technical fluency to run.
The commercial line formed over 2024 and 2025. Runway introduced Act-One on October 22, 2024 as a facial-performance tool — drive a character's face and dialogue from a simple recorded take, no rigging — and followed with Act-Two on July 15, 2025, extending the driving-video idea to head, face, body, and hand tracking. Kling's Motion Control (across its 2.6 and 3.0 releases) became the option most commonly recommended for full-body dance and action. ByteDance's DreamActor line pushed the pose-free approach, animating humans, cartoons, and animals from one reference image and one driving video, with its research accepted to a major graphics venue. Wan Animate (in the Wan 2.2 line) arrived as the open-source, self-hostable alternative. By 2026 the common thread across all of them was that a phone-recorded performance plus a single image was enough to animate a character — and that most of the head-to-head quality claims were vendor-reported.
| Platform | Behavior |
|---|---|
| Runway Act-One / Act-Two | The performance-and-expression option. Act-One (October 22, 2024) focuses on facial acting and dialogue delivery from a driving take; Act-Two (July 15, 2025) adds head, face, body, and hand tracking. Works from an ordinary phone or webcam recording and is usually positioned as the premium choice for dialogue and reaction shots. |
| Kling Motion Control | The one most recommended for full-body motion — dance, action, broad physical movement transferred from a reference clip onto a still character image and reused across characters or styles. Cited across 2.6 and 3.0 releases. Kling publishes internal benchmarks claiming wins over rivals; they are vendor-reported. |
| ByteDance DreamActor (M2.0) | The pose-free line. Animates any character — humans, cartoons, animals — from a single reference image plus one driving video, carrying motion and expression without an explicit skeleton. Best for non-human or heavily stylized targets. Published quality numbers are largely ByteDance's own. |
| Wan Animate (Wan 2.2) | The open-source, self-hostable option — customizable, cost-controlled, with a replace mode that preserves the source video's lighting. The control-and-budget pick rather than the top-of-the-leaderboard one; trades some polish for the ability to run and integrate the pipeline yourself. |
The trap I watch creators fall into with motion transfer is chasing the demo instead of the need. The viral examples are always a dance or an action shot mapped onto a wild character, so people assume that is the use case — and then they spend an afternoon making one impressive clip that has nothing to do with the recurring content their channel actually runs on. Be honest about what you need on screen every week. If the answer is "a consistent presenter saying this week's thing," motion transfer is the expensive, fiddly way to get there; a stable avatar or persona performance is almost always the more practical answer, and you reach for true motion transfer only when the specific performance — a particular dance, a stylized character acting out a bit — is the point of the piece.
The other half nobody budgets for is everything after the clip exists. A transferred shot is one asset; a channel is a weekly supply of on-brand pieces, cut to each platform, posted on a cadence. That production-and-distribution layer is where the time really goes, and it is where a full generation-and-publishing engine like [Kompozy](/) does the work motion-transfer models never touch — generating the recurring branded video with a consistent [avatar](/glossary/avatar-video), holding it all to one voice, and shipping the catalog, hero clip included, across every platform from one queue. Use the fancy model for the shot only it can make; do not let it become the whole plan.
It is a technique that takes the movement from one video and applies it to a different subject. You give the system a reference clip of something moving and a target — a photo or a generated character — and it produces new video of the target performing the reference's motion, with no motion-capture gear involved.
Motion capture uses sensor suits, markers, or depth cameras in a controlled setup and outputs rig-ready animation data an animator retargets onto a 3D model. Motion transfer reads movement from ordinary video with computer vision and outputs a finished rendered clip directly. Mocap gives more control; transfer gives more convenience and needs no hardware.
Motion transfer keeps the reference's motion and puts it on a new subject. Video-to-video restyling keeps the original clip's subject and motion and changes only the visual style. If the character in the output is different from the input, it is motion transfer; if it is the same subject in a new look, it is video-to-video.
Commonly cited ones in 2026 are Runway's Act-One and Act-Two, Kling's Motion Control, ByteDance's DreamActor, and the open-source Wan Animate. Act-Two leans to performance and expression, Kling to full-body motion, DreamActor to non-human and stylized characters, and Wan to self-hosted, customizable workflows.
Large, clearly visible whole-body movements — walking, dancing, broad gestures — transfer most reliably. Fine finger articulation, facial micro-expressions, and physical contact between multiple people are much harder and often produce artifacts, and a big mismatch between the reference body and the target shape strains the model. A clean, steady, single-subject driving clip matters most.