MiniMax H3 Recast: prepare a character replacement and check the result
Plan a MiniMax H3 Recast character replacement with source-video checks, reference-photo mapping, current fal limits and output review before a larger edit.

MiniMax H3 Recast refers to a video character-replacement workflow; fal lists its endpoint as H3 Max Recast. You supply an existing clip and reference photos for the new people. The provider describes preserving the source performance, camera movement, cuts and sound while replacing the cast. That is an intended behavior to check in the output, not a guarantee that every frame will remain unchanged.
This guide uses the official fal model page and input documentation, checked on October 11, 2026. We did not submit a Recast generation. The example below is an original preparation brief, not a tested result. Yix does not currently offer this specific Recast workflow.
Cover: an Imagen-generated illustration of matching a character reference to a film sequence. It is not a Recast output or a screenshot of the provider's interface.
Separate Recast from reference-to-video
A reference-to-video model uses supplied media to guide a new generation. Recast begins with a source sequence whose timing and performance you want to retain. Those are different tasks even when the model family names overlap. Choose by the required input and output behavior, not only by a label containing “H3.”
For example, “make a new scene starring this character” asks for a new sequence. “Replace the person in this existing shot while keeping their gesture and the camera move” describes the Recast task. If you mainly need a new action or a different shot layout, decide that before spending time preparing a source-preserving replacement.
Prepare the source clip and cast map
The documented input requirements accept a 5–30 second source video, with no individual shot longer than 15 seconds, and one to four reference photos. Each photo represents one replacement person. The default mapping follows the main people from left to right; an optional prompt can clarify who should replace whom.
The total duration limit and the shot limit are separate. A 20-second file containing two 10-second shots is a different input from one uninterrupted 20-second shot. Inspect actual cut points, not just the file's duration. A simple first trial can use one short shot with one clearly visible adult character and a stable reference photo.
Use footage and portraits you have permission to edit and publish. Favor a reference with an unobstructed face, clear lighting and enough detail to inspect. Avoid introducing several people in one reference photo when the task calls for one new person per image. A clean input will not guarantee success, but it makes failures easier to understand.
Before uploading, write a cast map outside the prompt:
| Source role | Replacement reference | Detail to inspect afterwards |
|---|---|---|
| Main character in the foreground | Reference A | Face through turns and changes of expression |
| A second character, if needed | Reference B | No identity switch when people cross or overlap |
| Background people not intended to change | No assigned replacement | Whether the output changed them unexpectedly |
Treat this as your inspection plan. The provider's person-tracking description does not establish that every occlusion or crowded shot will be handled correctly. For a first evaluation, simplify the source before adding a more difficult sequence.
An original brief for a small test
Imagine a fictional costume scene: one adult in a red jacket turns toward the camera and lifts a paper lantern. You want a new fictional character from Reference A while retaining the gesture, framing and scene. A concise, untested brief could be:
Replace the foreground adult in the red jacket with the adult character in Reference A. Keep the source action, jacket, paper lantern, camera movement, shot timing and background. Do not add new people or change the scene.
Use that as a structure, not a guaranteed recipe. Name the role and the intended changes, then state the details that matter. If the source has several people, describe the target clearly enough that you can later assess whether the mapping was followed. Avoid packing unrelated scene changes into a trial whose purpose is to evaluate the replacement.
Check resolution and the provider's cost
As checked on October 11, fal lists 768p at $0.30 per output second and 1080p at $0.45 per output second, with reference images included. These are fal's displayed rates, not Yix credits or a quote from another platform. The documented default resolution is 1080p. Confirm the current estimate before submitting.
| Output duration used for the estimate | 768p | 1080p |
|---|---|---|
| 5 seconds | $1.50 | $2.25 |
| 10 seconds | $3.00 | $4.50 |
| 30 seconds | $9.00 | $13.50 |
These figures are simple duration-times-rate examples. They do not promise that a particular source will succeed or describe any account-specific billing adjustments. Repeated generations add cost, so decide what you are testing before changing the prompt or reference.
The input controls also include a realism option and a random seed. Start with the controls needed for your task and note the settings you use. If you compare two results, change one meaningful input at a time; otherwise it becomes difficult to tell whether a difference came from the reference, wording, resolution or random variation.
Review the saved output, not just a poster frame
Watch the full clip once at normal speed, then inspect difficult moments around turns, hand-to-face contact, overlaps and cuts. A strong first frame can hide drift later. Compare against the original clip so you can identify changes outside the intended cast replacement.
Check face identity, expression, hair edges and skin transitions without treating them as separate from the performance. Then inspect clothing, props and hands. The new character may look plausible while an important object changes shape or a gesture becomes less clear.
Listen to the saved file as well. The documented output includes source sound; verify timing and continuity instead of assuming the soundtrack survived correctly. Confirm duration, framing and resolution in the actual downloaded file. Keep the source clip, reference order, prompt and settings with the result so a later revision has a reproducible starting point.
Fit the result into the wider edit
If the replacement passes your checks, bring the downloaded clip into your editing project and review the neighboring shots. Consistent identity inside one short output does not prove continuity with the rest of a film. Color, wardrobe, screen direction and audio transitions still need editorial judgment.
Our FilmCraft guide covers planning a timeline and checking a small export. If you are still creating source footage, the Yix video workspace serves that separate stage; it is not a shortcut to H3 Max Recast. Keep those steps distinct so you choose the tool for the change you actually need.