How to replace a person in a video with AI
Prepare a short reference video and a character photo, write a clear reference prompt, review the quote and inspect a character-led video remake.
Choose one clear shot
Begin with a short continuous clip containing one clearly visible person. A steady camera and a simple action make it easier to see whether the new character remains consistent. Fast cuts, crowds and heavy occlusion combine several problems in the same test.
Use a video you own or have permission to adapt, and obtain permission from the person whose image you use. A public social-media link can help you identify a style, but it does not supply a source file or permission. CastRemix takes uploaded files; it does not download clips from social platforms.
Match the character photo to the shot
Use a sharp, well-lit photo. For a standing or dancing reference, include the full body. For a presenter, include shoulders and any hands that matter to the shot. A very tight face crop leaves the model to infer clothing, pose and body proportions.
Start with one character image. Add another image only when it contributes useful information, such as a clearer outfit or a product that needs its own reference. More references can introduce conflicting identities rather than solve them.
Describe the role of each reference
In the composer, add the photo and video, then click their labels or type @ to insert references. A useful starting instruction is: “Use @Image1 as the character identity. Recreate the simple movement and framing in @Video1. Keep one performer and consistent clothing.”
Describe what the photo contributes and what the video contributes separately. Avoid asking for a new location, new camera move and new choreography in the same first test. This makes it easier to attribute a change to a specific instruction.
Review output settings and credits
The workspace shows the current model for reference-led generation. Choose aspect ratio, duration and resolution, then check the displayed quote. Reference-video duration contributes to the cost alongside output duration. Uploaded images do not add timed charges.
If generation is not connected, you can still prepare the draft; the workspace will show its availability state. Do not treat a prepared prompt or a concept illustration as a completed video. For a low-cost first check, use one short reference and the shortest suitable output.
Inspect the whole clip before sharing
Compare the opening, middle and final frames for identity and clothing. Then watch hands, body proportions, floor contact and interactions with objects. Review background and camera changes as separate checks. A good opening frame does not establish that the entire clip is usable.
Reference-led generation is a remake, not a promise of frame-exact editing. Exact face replacement, an unchanged background and precise lip sync are not guaranteed by this workflow. Keep a useful result as a draft and add final audio, captions and disclosures in your editing process.