Features
AI Motion Transfer Video Generator
By Ofrai Editorial (Use Case Research)
AI motion transfer takes the movement from one video and applies it to a character from another source, so a performance you already have can play out as your own character without re-shooting it. You supply a driving clip for the timing and a character image for the appearance, and Ofrai's video engines re-time the body frame by frame. That makes it practical for choreography, sports drills, martial arts and product motion where the move matters more than the face.

Chat and video use credits from your Ofrai balance.
Example prompts
Cinematic chase, handheld tracking, rainy street, neon streaks, accelerating pace
What motion transfer actually moves
Motion transfer is a two-input job: a driving clip that supplies timing, pose and velocity, and a character image that supplies appearance. The engine does not invent choreography — it re-times your character's body to match the driving performance frame by frame. That distinction matters for prompt writing. Describe what the character is doing inside the shot, not how they should move; describing the move competes with the reference and gives you a body that half-follows the instruction and half-follows the clip.
Choosing a driving clip that transfers cleanly
Clip quality matters more than clip length. Choose a driving video where the full body stays visible for most of the take, the background is simple, and the subject is not heavily occluded. Cuts, whip pans and motion blur all break the pose estimate between frames, and the artefact appears as limbs snapping or feet sliding. A static tripod shot of a full-body performance transfers far better than a handsome handheld clip. Ten usable seconds beat forty seconds containing two cuts.
Preparing the character image before it moves
The character image is a pose problem. Use a frontal, full-body frame with feet visible and arms clear of the torso, so the model has a clean silhouette to articulate. A waist-up portrait gives the motion model nothing to work with below the hips and it will invent legs. Choose even lighting with no hard cast shadows, avoid extreme perspective, and start from a neutral stance. If identity must hold through the clip, add a second clean reference of the face alongside the body frame.
Camera notes that keep a transfer believable
Camera notes decide whether the result looks shot or looks assembled. State shot size, angle and movement explicitly: medium full shot, eye level, locked-off tripod. If the driving clip has camera motion, the character inherits it, so ask for it rather than hoping for a static result. Keep lens language consistent with the character image — a 24mm wide look on a figure drawn at 85mm compresses proportions and starts reading as a different person partway through.
Ratio, duration, and where to cut
Render at the aspect ratio of the driving clip rather than the still, because the pose geometry arrives pre-composed. For social delivery, 9:16 throughout is the simplest route. Match duration to the driving take: a gentle push-in on a short clip is fine, but asking the engine to invent choreography after the reference ends is where quality collapses. If you need 20 seconds, transfer two 8-second segments and cut between them in an editor where motion hides the seam.
Why motion transfer smears, and how to fix it
Four failure modes. Limb snapping at fast direction changes: choose a driving clip with smoother transitions. Feet sliding: the character's footwear does not match the ground plane, so pick a flat, evenly lit floor in the character image. Costume merging: long coats, skirts and loose hair get pulled toward the driving silhouette, so simplify the outfit before transferring. Late identity drift: re-anchor by re-uploading the original character image at the start of every segment.
Choosing an engine for reference-driven motion
kling-3 is the strongest of these four on complex human articulation and choreography, which is most of what transfer consists of. kling-omni is useful when the driving clip and the character come from different sources and the engine has to reconcile them, or when the performer interacts with props. seedance-2-5 delivers fast, clean results on simpler repetitive movement. wan-2-6 is the cost-first pass for testing timing and framing before spending on a final render.
Three motion transfer prompts with the settings spelled out
First, a walking turn: character image plus driving clip, prompt "Walk three steps, stop, turn to camera left over two seconds, hold the final pose. Keep the original face, outfit and background. Static camera." Second, a gesture with props: "Lift a ceramic cup to the mouth and set it back down, elbow close to the body. Preserve the original clothing and the room exactly." Third, a repeated action: "Repeat the same arm-circle twice at a steady tempo, weight shifting through the hips, no camera movement."
In each case the motion comes from the driving clip and the prompt only describes what the transfer should do with it. Adding a camera instruction on top of a driving clip that already has camera motion is the single most common cause of a doubled, unusable take.
Driving video and character image: who owns what
Split the authority between the two inputs or they will fight. The driving video owns timing, body path, weight shifts, gesture amplitude and the rhythm of the movement. The character image owns identity, face, hair, wardrobe, body proportions and the environment. Everything the two overlap on — pose at frame zero, apparent body shape, apparent lighting — is where drift appears, which is why the best transfer jobs are ones where the performer in the driving clip is wearing something roughly similar to the character.
A tailored jacket in both, or loose athletic wear in both, transfers cleanly; a floor-length coat on the character against a vest-and-shorts performer will produce smearing around the hem on every frame. Match the wardrobe category, not the exact garment, and keep the camera height of the driving clip consistent with the framing of the character image so the ground plane agrees.
Timing the transfer so hands and feet land on the beat
Where the action starts in the driving clip sets where it starts in the output, and that is the control you have over timing. If the transfer starts one beat late, trim the head of the driving clip rather than writing "after a delay" in the prompt — trimming is exact and prompting is approximate. Keep the first half-second of the driving clip relatively static when the action involves a precise landing, so the engine has an unambiguous rest pose to depart from.
For repeated actions, choose driving clips built on a countable cycle — two steps, one reach, three claps — because the model reproduces a counted cycle far more faithfully than an uncounted phrase. And if the result is one beat off the music underneath, fix it in the edit with a two-frame trim rather than regenerating: the movement will be correct, only the sync is off, and that is the cheapest fix available in video work.
Credit and render-time expectations per take
Budget per take rather than per finished clip, because transfer produces a lot of near-misses. A single action at five seconds on one of the engines listed on this page typically needs two to four attempts: one to establish whether the transfer works at all with your inputs, one or two to fix the specific artefact, and occasionally a fourth for a clean final.
Longer driving clips cost proportionally more and drift more, so a six-second target is usually cheaper than a ten-second one even when you need the longer clip, because you will regenerate the ten-second version more often. Use wan-2-6 to confirm timing and framing at low cost, then run only the takes you have already validated on kling-3 or seedance-2-5. If a driving clip keeps failing, the cheapest fix is a different clip, not a fourth attempt on the same one.
Motion transfer versus dance transfer and outfit swap
All three read two inputs, but they read different things and they fail differently. Motion transfer copies the movement of a driving clip onto your character without caring about beat, count or musical phrasing; that makes it the right tool for a walk, a turn, a gesture, a stretch, or a mascot performance where the timing is yours.
Dance transfer is the same mechanism with a stricter contract: the beat, the count and the phrase length all have to line up with the music, so it needs a driving clip that was cut to the track. Outfit swap keeps the motion from your own footage and replaces the wardrobe instead, which means it is not a transfer job at all and the driving clip is your original shoot. If your reference video is music, use the dance page.
If it is footage of someone else performing, use this page. If you want new clothes on footage you shot, use the outfit swap page.
FAQ
- What is the difference between motion transfer and motion control?
- Motion control is a rig: you drive specific joints or a camera path with parameters. Motion transfer is closer to performance retargeting — the engine reads a driving video and makes your character match its timing, pose and velocity throughout. For a full-body dance or fight you filmed, transfer is right; for a locked camera move around a static character, control is cheaper and more predictable.
- Does the driving clip need to show a face?
- No. A driving clip is read for skeleton, timing and velocity, so a rehearsal, a side view, or loose clothing often transfers more cleanly than a polished performance. What matters is that limbs stay visible, the background is simple, and the frame rate is steady. Keep the character image separate and high quality.
- Why does the character drift after a few seconds?
- Identity drift accumulates because the model re-samples the subject on every frame it generates. Keep the character image as a constant reference, generate shorter segments and cut them together, and re-upload the original reference at the start of each segment. Long single-pass generations of 15 seconds or more are where drift becomes obvious.
- Can motion transfer work with a full-body photo?
- Yes, as long as the body is fully visible: head to feet, arms clear of the torso, feet in contact with a visible floor. Waist-up images leave the motion model nothing below the hips and it invents legs. If only a portrait exists, generate a full-body still in Image Studio first, then feed that frame into the transfer.
- Which Ofrai engine is best for motion transfer?
- kling-3 for the most reliable human articulation and complex choreography, kling-omni when the driving clip and character come from different sources or the scene includes props, seedance-2-5 for fast clean results on simpler movement, and wan-2-6 when you are testing timing on a budget. All four draw on the same Ofrai credits.
Related generators
- AI Dance Transfer For Choreography Reuse
- AI Object Replacement Video
- AI Outfit Swap Video
- AI Photo Dance Video Generator
- AI Product Swap Video
- AI Virtual Try-On Video
- Seedance 1.5 Pro on Ofrai
- Photo to video
- AI Animal Video Generator
- AI Anime Video Generator
- AI Furry Video Generator
- AI Realistic Video Generator