How AI recasting works: your photos, one reference clip, a new video

"Recasting" means taking an existing clip and replacing the people in it while the rest of the scene stays put. A few years ago that took a VFX team. Today a video model can do it from one photo per person.

Two references, one instruction

A recast starts with two inputs:

• The reference clip: the original take, which defines the location, camera movement, timing, crowd and sound.
• The reference photos: one per person who should take a role, such as the lead singer or the sax player.

The model also gets a written instruction: recreate the clip shot for shot, put the person in each photo in the place of one named performer, and keep everyone and everything else unchanged.

What the model is good at

Modern reference-to-video models, such as ByteDance's Seedance, are trained to keep a character consistent across a whole clip. They can follow the motion of a reference video closely while swapping in the identity from an image. They can also generate audio, or carry a reference soundtrack through.

Where it still struggles

• Hands and fast motion: quick gestures can blur or gain extra fingers.
• Profile angles: if the photo only shows the front, side views are partly guessed.
• Text and small details: signs and logos in the background can warp.
• Long clips and big casts: the longer the take and the more people you swap, the more chance a likeness drifts. Strong input photos matter more here.

Why it costs money per video

Each render runs a large model on GPUs for several minutes, and the provider bills per second of video, counting both the reference clip and the output. That's why recasts are sold per video rather than as an unlimited plan, and why a failed render gets its credit back.

Use it kindly

Recasting is a creative tool, and it can also be misused. Only recast people who agreed to it, never minors, and label the result as AI-generated when you share it.