MiniMax H3 Recast Reference Photos: How to Map 1–4 People

MiniMax H3 Recast accepts 1–4 reference photos, with one photo for each new person. Their order matters: by default, the endpoint maps the photos to the main people in the source video from left to right, then follows those people through the shots in which they appear.

This guide is only about preparing and checking that person map. Read the MiniMax H3 Recast overview first if you still need to decide whether replacing people in an existing video is the right operation. This site does not run Recast or accept a source-video upload.

Use one person per reference photo

The current fal Recast schema requires at least one reference image and permits no more than four. Give each replacement person a separate image rather than a group photo.

Choose a photo in which the intended person is easy to identify. A visible, unobstructed face, even lighting, and enough detail to inspect the eyes, mouth, hairline, and jaw make the mapping easier to review. Crop out unrelated people instead of leaving the endpoint to infer which face matters.

A clear portrait is useful input, not a guarantee. No pose, crop, lighting setup, or prompt can promise an exact face, hairstyle, garment, body shape, or lip movement in every output frame.

Write the map before uploading anything

For every main person in the source, record the source role and the replacement reference. Use stable details from the source video rather than relying only on a screen position that may change after a cut.

Upload the photos in that same order. In a simple two-shot, Reference 1 maps to the main person on the left and Reference 2 maps to the main person on the right by default.

Do not add a fifth photo. If more than four important people need new identities, split the work into separately reviewable source clips or use a workflow whose published contract supports that cast size.

Treat crossings and cuts as mapping risks

Left-to-right order is easy to inspect in a stable wide shot. It becomes less descriptive when people cross, leave the frame, re-enter, or appear alone after a cut.

Imagine a café sequence:

1. the barista starts on the left and the customer sits on the right;
2. the customer appears alone in close-up;
3. the barista walks around the counter and finishes on the right.

“Left person” is no longer a useful description by the third shot. Keep the photo order fixed, then identify each source person by a role, wardrobe detail, or continuing action that survives the edit.

The schema says the default mapping follows each main person in every shot where they appear. That describes the intended mapping behavior, not a guarantee that every identity will remain perfect through every crossing, occlusion, angle, or cut. Review the complete output.

Use the prompt to name who becomes whom

The optional prompt can disambiguate the map. Keep it within the published 2,000-character limit and focus on visible source roles:

Replace the barista wearing the green apron with Reference 1 in every shot.
Replace the seated customer wearing the cream sweater with Reference 2.
Keep the source clothing, hand movements, camera movement, cuts, timing,
and sound.

The clothing descriptions identify people in the source; they do not ask the people in the reference photos to bring their portrait wardrobes into the video. If a wardrobe change is essential, state it separately and treat the result as something to verify, not something the prompt guarantees.

Avoid labels such as “person one” when the scene itself offers a clearer anchor. A useful role description answers two questions: which source person is this, and which numbered reference should replace them?

Confirm consent on both sides of the map

The source performer and the person in the replacement photo are different rights and consent checks. Before submitting a job, confirm that you are allowed to transform the source performance and to use every reference person's likeness for the intended context and distribution.

If the result could make viewers believe a real person performed an action or delivered dialogue they did not perform, disclose that the video was recast. A provider's commercial-use label does not by itself grant rights to a source video, a face, a voice, a brand, or a final use.

Review the mapping shot by shot

Keep the written map beside the result and inspect every shot in order:

1. Opening map: confirm that each numbered reference starts on the intended source person.
2. Solo shots: confirm that a close-up still belongs to the role assigned in the wider shot.
3. Crossings: follow the person, not their new left-or-right position.
4. Cuts and re-entry: check the first frames after each edit and whenever someone returns.
5. Final identity pass: compare each replacement with its intended reference across the whole clip, not only the cleanest frame.

If two identities are reversed, correct the reference order or make the role prompt more explicit before changing unrelated settings. A new seed is not a substitute for an unambiguous map.

Developers implementing these fields can continue with the MiniMax H3 Recast API guide, which covers server-side credentials, queued jobs, public input URLs, and result ownership.