How to Make an AI Motion Transfer Video from a Dance Reference

To make an AI motion transfer video from a dance reference, start with a readable movement source and a separate image for the new subject. ArtArch's agent workflow checks the dance reference, prepares a depth motion guide, and uses that guide to drive the subject image through the same body movement. The ArtArch CLI page is the starting point for installing the skill.

Choose a reference that a beginner can read

The first result depends on the reference more than on a long description. Look for a clip with one main performer, visible body movement, a stable aspect ratio, and enough space around the hands and feet. A clean full-body dance is easier to transfer than a cropped close-up.

Before any model processing, the action-reenactment skill extracts compressed representative frames. The agent uses this inspection to confirm that the video has a decodable video stream, that human movement is the main subject, and that the number of simultaneous people is understood.

Fast motion blur, strong occlusion, and hard cuts deserve attention at this stage. They can interrupt a body track or reset the visual relationship between one segment and the next. Reviewing the frames first gives you a chance to choose a better source.

Use two references with different jobs

The subject image owns the new character. It controls appearance, proportions, face, wardrobe, and stable accessories. A full-body image makes it easier for the final clip to show the movement you care about.

The tracked-depth video owns the selected motion. It carries action order, timing, body silhouette, screen position, and weight shifts. The original dance video remains the source used to prepare this guide.

If the final shot needs the same setting, an environment frame can provide the location, viewpoint, composition, and lighting. People, faces, wardrobe, stickers, and foreground subjects in that frame are not part of the new subject's identity.

The final generation pattern uses one tracked-depth video reference. Keeping the original dance video out of that step helps separate movement from the source performer's appearance and background action.

A simple four-step workflow

1. Give your agent a character image and a dance reference video. State the action you want the new character to follow.
2. Let the agent inspect compressed frames and report the visible performer count, cuts, blur, and occlusion.
3. Review the grayscale depth guide and colored person-track preview. Confirm that the selected dancer's movement is the motion you want.
4. Approve the final generation after the agent shows the connected subject image, optional environment frame, and single tracked-depth video reference.

The ArtArch CLI page demonstrates this kind of request:

Make the character in this image perform the dance from the reference video.