Copy Motion From a Reference Video: Which AI Platform for What

Describing motion in words is hard. "She dances" can mean anything; a ten second clip of the actual dance means one thing. Motion transfer lets you act out a movement, or borrow it from footage you have rights to, and apply it to a generated character. In 2026 this split into two families: dedicated performance tools that drive one character from one video, and reference models that take motion as one input among many. This guide covers both, as of November 2026.

Motion transfer at a glance

Performance tools

Runway Act-Two

Act-Two takes a driving performance video, someone acting the scene, and a character image or video, and moves the character to match. With a character image, you can let the performer's hands and body drive the character's gestures; turn gestures off and Act-Two focuses on facial and environmental motion. Runway's tips: keep the performer's hands in frame at the start, begin in a pose close to the character's, and favour natural movement over abrupt changes. It needs a Standard plan or higher.

Kling Motion Control

Kling Motion Control drives one character in an image with movement from a video you upload, or from its motion library. The latest version is built on Kling VIDEO 3.0, with better facial consistency than 2.6. Motion videos must be 3 to 30 seconds and the output matches their length. Two orientation settings matter: matches video copies movement, expressions, camera movement and facing from the reference, while matches image keeps the character's facing from your image and lets you prompt the camera. You can also bind a face reference (an element) to keep the face consistent. Kling advises one continuous shot with one character, full body and head visible, and no cuts.

Luma Ray3 Modify

Luma's Ray3 Modify (December 2025) is a video-to-video model for actors: it keeps the original motion, timing, eye line and emotional delivery of your footage and maps the performer onto a character from an image you provide. Start and end frames let you direct transitions. It runs inside Luma's Dream Machine app.

Reference models

Hailuo H3

MiniMax's H3 Omni Reference mode accepts up to 9 images, up to 3 video clips and up to 3 audio clips, 12 files in total, with video and audio clips of 2 to 15 seconds each and 15 seconds combined. That lets you give a character image, a motion clip and a voice sample in one generation, with native stereo sound. MiniMax open-sourced H3 in August 2026, so it also runs outside its own apps.

Seedance 2.5

Seedance 2.5 takes up to 50 references, including motion clips, and supports white-model references: rough 3D blocking of where things move, which the model dresses into a finished scene. Useful when you want to block motion yourself rather than film it. See our Dreamina guide for the app it lives in.

Wan 3.0

Alibaba describes Wan 3.0 (August 2026) as an all-in-one reference model covering generation, editing, visual replication and character driving, with clips up to 30 seconds. It is API-first, through Alibaba Cloud and partner platforms.

Getting good results

• Match the framing. A full-body dance onto a head-and-shoulders portrait fails. Use a character image framed like the reference.
• One subject, one continuous shot. Cuts and extra people confuse every tool.
• Plain background, steady camera in the reference, unless you want the camera copied too.
• Moderate speed. Very fast moves come out blurred or shortened; Kling notes complex motion can produce a shorter output.
• Use footage you have the right to use, such as your own performance or licensed material.

To keep the character itself consistent across several of these shots, see consistent characters: which platform for what.

Try it

Film yourself for ten seconds on a phone against a plain wall: walk two steps forward, wave, and say one line. Use it as the driving video with one character still in two tools, adding this prompt where the tool takes one.

A friendly cartoon fox shopkeeper behind a wooden counter in a cosy village shop, warm afternoon light, shelves of jars behind him. He steps forward, waves and says "Morning! What can I get you?" Keep the camera static at counter height.

Compare the timing of the wave, whether the line lands on the lips, and whether the fox stays the same fox throughout.

Motion work produces a pile of takes, each tied to a different reference clip. Dreamdrive, a free Chrome extension, saves every result as you make it with its prompt, model and settings in one library, so you can find the take that worked and see how it was made.