A shot contract for AI character replacement: what must stay, what may change

When you ask an AI video tool to “swap the character,” it may produce a good-looking different shot. That can be the wrong result. Before generating, write a short shot contract: what the source video must control, what the character image may change, and what variation you will tolerate.

This is a planning brief, not a magic prompt. Some tools have a prompt box; others, including the browser workflow I built, use the video and image directly. The contract helps you choose inputs and judge the result in either case.

Copy this shot contract

Goal: Replace ONE main visible subject in an authorised reference clip.

MUST KEEP from the video:
- The order and timing of the action: [describe the key beat]
- Camera position/movement and framing: [describe]
- Background and important props: [describe]
- Original audio: [keep / mute; verify that you may use it]

MAY CHANGE from the character image:
- The main subject's face, clothing and overall appearance

ACCEPTABLE VARIATION:
- Small generative differences in hair, fabric or fine background detail

FAIL IF:
- A second person changes unexpectedly
- The main gesture, camera move or shot order is rewritten
- Audio starts late, disappears, or no longer matches the visible beat
- The replacement identity drifts during a turn or occlusion

Fill the brackets with observable details. “Cinematic and consistent” is too vague to review. “The subject raises the cup at about two seconds while the camera stays locked” gives you a concrete comparison.

A five-second example

Suppose your own clip shows one person turning toward the camera and lifting a cup. You want your original illustrated character in their place. The person, cup, background, and camera do not all have the same role:

• The video defines the turn, cup lift, timing, camera and room.
• The image defines who performs that movement.
• If the original dialogue matters, choose to keep the source sound. If not, make a silent result and add approved audio later.

Choose a reference with one clear main subject. A two-person conversation is not automatically a two-character replacement request; for a single-subject workflow, it is an ambiguous reference. Likewise, a still portrait with a radically different head angle from the visible subject makes identity harder to carry through the turn.

Review it as an edit, not just an image

Use a simple four-point check:

1. First frame: Does the intended character appear in the right place and scale?
2. Hardest moment: Does identity survive a profile turn, motion blur, or an object crossing the face?
3. Last frame: Is it still the same character, with the intended action complete?
4. Real-time playback: Are gestures, cuts and—if retained—sound still aligned?

If the character changes but the gesture is also rewritten, choose a clearer reference beat or split a complicated edit into shorter shots. If the motion remains but the replacement identity drifts, try a sharper character image with a more compatible angle. If sound is wrong, diagnose that separately; do not assume a visual replacement tool also rewrites or repairs dialogue.

The point of the contract is to change one variable per retry. Otherwise, you cannot tell whether the next result improved because of the image, the clip, or the audio setting.

Keep the scope honest

This method is best suited to an authorised short clip with one main subject. It is not a promise that every person in a crowded scene can be independently replaced, nor that a generative model will preserve every pixel. Check rights to the footage, character and soundtrack before publishing.

Disclosure: I built UseLipSync. Its Recreate workflow accepts your own 4–15 second MP4/MOV reference and one authorised character image for a single main subject; it can retain the original audio when selected, and the result must be reviewed before use. The step-by-step example with real source/output footage shows what this input contract looks like in practice.