Minimax H3 Recast

It performs a cast substitution rather than generating video from nothing: the portraits you supply take the place of the people already in your footage, while movement, framing, edits and audio all carry over from the original.

Your clip should fall between 5 and 30 seconds, with no single shot longer than 15 seconds. Split longer takes before uploading, because the output is returned at the same length as the input.

Up to four, each with its own portrait. By default they map to the leads in screen order, so arrange the uploads accordingly; a written instruction can override that mapping.

Use one sharp, evenly lit, front-facing portrait per person. Half-body framing carries hair and clothing over more faithfully than a close head crop; avoid sunglasses, hats, hidden faces, group photos, heavy filters and blurry images.

Run the low-resolution pass first to check the pairing, then produce the final file in 1080p. Since cost follows clip length, that first pass is the cheap way to catch a mismatch.

It will not help when the footage itself must change — new movement, new camera work, a new location — because those come entirely from your source. Results also weaken with dense crowds, distant or tiny faces, fast motion, and anything blocking the face after a hard cut; across long takes, or when someone turns to profile, the likeness can drift back toward the original performer.