Motion transfer takes the movement in one clip and drives something else with it. You give it a reference video and a target, a character image or an object, and the result is your target performing the reference's motion. It is one of the more reliable ways to get natural movement out of AI video, because the movement is captured from real footage rather than invented from a prompt. It also has clear limits. This guide covers what transfers cleanly, what tends to break, and how to prepare a reference so the result holds.
What motion transfer actually does
The system reads the reference clip frame by frame and extracts the movement in it: pose, gesture, timing, the arc of a limb, the rhythm of a walk. It then maps that movement onto your target while trying to keep the target's own identity intact, the face, the proportions, the art style. The output is a new clip that borrows the reference's motion and keeps your subject's look.
The reason this is worth reaching for is that motion is the hardest thing to describe in words. You can write "she walks confidently" and get ten different walks. A reference clip removes the ambiguity: the model is copying an actual movement instead of guessing at an adjective. That is why motion transfer often looks more natural than a pure text-to-video generation of the same action.
What holds up
Some things transfer reliably. Plan around these and the result tends to land on the first try.
- Whole-body, clearly visible motion. Walking, dancing, gesturing, a sports move performed in full frame. When the body is unobstructed and the movement is large, the mapping has plenty to read.
- A steady rhythm. Repeating or continuous motion, a sway, a stride, a wave, gives the model a consistent signal and produces the smoothest output.
- Motion that suits the target's build. A reference whose proportions are close to your character's transfers more cleanly than one that is wildly different, because the model has less to reinterpret.
What tends to break
Other cases are where results wobble. None of these are hard stops, but they are worth knowing before you spend a generation on them.
- Occlusion. When the subject in the reference is hidden behind furniture, cropped at the edges, or turned away, the model loses the part of the pose it cannot see and has to guess.
- Fast, tiny detail. Individual fingers, quick hand articulation, and subtle facial motion are the first things to distort. The larger the motion, the safer it is.
- Extreme mismatch of form. Driving a four-legged animal with a two-legged dance, or a tiny object with a full human gait, asks the model to reconcile shapes that do not correspond, and the result shows the strain.
- Long takes. As with any AI video, the further into the clip you go, the more the target's identity can drift. Short references hold better than long ones.
How to prep a clean reference
1.
Frame the whole subject
Pick or shoot a reference where the body performing the motion is fully in frame from start to finish. If limbs leave the frame or the subject is partly hidden, the model has to invent the missing pose, which is where distortion starts. A front-facing, well-lit subject on a plain background is the cleanest input you can give.
2.
Keep it short and continuous
A reference of a few seconds carrying one clear movement beats a long clip with several actions stitched together. Trim to the beat you actually want transferred. If you need two distinct movements, run them as two passes rather than asking one clip to carry both.
3.
Match the target to the motion
Choose a character or object whose proportions are in the same neighbourhood as the reference subject. A standing, front-facing character image gives the model the cleanest starting pose to move from. The closer the forms, the less the model has to reinterpret and the more the identity survives.
4.
Generate, then judge the two things separately
When the clip comes back, check motion accuracy and identity preservation as two separate questions. Did the movement transfer with the right timing? And does the character still look like itself throughout, not just in the first frame? If one is off, you usually fix it by changing one input: a cleaner reference for motion, a stronger target image for identity.
When to reach for it, and when not to
Motion transfer is the right tool when you have a movement you want exactly, a dance, a specific gesture, a captured performance, and a character or object you want to perform it. It is the wrong tool when the motion is incidental and only the look matters, in which case a plain image-to-video generation is simpler. It is also not a substitute for video-to-video restyling, which changes the look of a clip while keeping its motion, the opposite of what transfer does. Knowing which of the three you actually want is most of the battle.
Prep the reference well and motion transfer is one of the more dependable corners of AI video. You can save a version you like as a repeatable step in the Morphic Workflows library and run it against new characters whenever you need the same movement again.
FAQs
Motion transfer reads the movement in a reference video, pose, gesture, and timing, and applies it to a different target, usually a character image or an object. The result is your target performing the reference clip's motion while keeping its own look. Because the movement is captured from real footage rather than described in words, it often looks more natural than generating the same action from a text prompt.
A short clip, a few seconds, where the subject performing the motion is fully in frame, well lit, and against a plain background. Whole-body, clearly visible movement like walking, dancing, or gesturing transfers cleanly. Avoid references where the subject is cropped, occluded, or turned away, since the model has to guess at the parts of the pose it cannot see.
Fine, fast detail is the first thing to distort: individual fingers, quick hand articulation, and subtle facial motion. Larger movements are safer than small ones. Identity can also drift over a long clip, so keeping the reference short and choosing a target whose proportions are close to the reference subject both help the character hold together.
Yes, as long as the target's form is compatible with the reference motion. Driving a shape with movement that suits it works well; asking a very different form to follow a mismatched gait, like a small object performing a human stride, forces the model to reconcile shapes that do not correspond and the strain shows. Match the motion to the target's build.
They do opposite things. Motion transfer keeps the reference's motion and puts it on a new subject. Video-to-video restyling keeps the original clip's motion and changes its look. If you want a specific movement performed by your character, use motion transfer. If you want to restyle footage you already have while its motion stays put, use video-to-video.
Yes. Run the process once per character using the same reference clip. Each run produces an independent result of that character performing the reference motion, which is a fast way to build a set of variations, for example the same gesture across a cast, without re-sourcing the movement each time.