MiniMax H3 Recast is a search term for replacing people in an existing video while keeping the shot's performance and structure. Start with a clear editorial question: do you want to change who appears in a finished clip, or create a new clip around a reference character? Those jobs need different inputs and different acceptance checks.
What MiniMax H3 Recast does
The fal Recast documentation describes a source-video workflow with reference photos for the replacement people. Its documented intent includes retaining motion, camera work, cuts and source audio. The published schema accepts a 5–30-second source clip, with no individual shot longer than 15 seconds, and offers 768P or 1080P output. These are fal endpoint details checked on October 9, 2026, rather than specifications for every H3 interface.
This guide focuses on preparing and evaluating that kind of edit. If your task is a new scene driven by character references, read the H3 Max Ref planning guide. For a new video from an idea or a starting image, create an H3 video and build the scene with the controls available in your chosen generator.
Prepare a MiniMax H3 Recast source shot
Choose a clip where the intended subject can be identified without guessing. A medium shot with stable lighting is a useful first experiment. A fast turn, a hand crossing the face, a mirror reflection and a second person entering the foreground each add a separate thing you must inspect. Do not combine all of them in your first test.
Make a shot inventory before uploading anything. For each cut, record the person's position, outfit, entrance and exit, and any moment when another object covers the face. The inventory becomes your review checklist. If a source has three cuts, inspect the first frame after each cut as carefully as the opening frame; a good beginning does not establish consistency throughout the edit.
Use reference photographs that you have permission to use. Pick a clear face with enough detail to compare eyes, nose, hairline and expression. Avoid treating a heavily filtered portrait as an exact identity target. When several replacements are involved, label the files by their intended roles and keep those labels in your edit notes.
Write an explicit cast mapping
The provider documentation describes left-to-right replacement as the default mapping and an optional prompt for directing the edit. State the mapping in plain words when the arrangement is ambiguous. Our example is an original editorial brief, not a tested model output:
Replace the presenter nearest the window with the person in reference photo 1.
Keep the second presenter, desk, wardrobe colors and background unchanged.
Preserve the source shot order and the presenter's original gesture timing.
A mapping should name the target before describing aesthetic changes. If you also ask for new lighting, a different lens, a new outfit and a redesigned set, it becomes harder to diagnose why a result differs from the source. Run the casting experiment first, then decide whether the creative direction needs another pass.
Review identity, motion and continuity separately
Watch the result once at normal speed without pausing. Write down any moment where the face feels unstable or the performance looks interrupted. Then compare still frames at the opening, a turn, an occlusion and the end. Review at delivery size as well as zoomed in; an edit that looks acceptable on a phone may show obvious boundaries on a large display.
Check body and scene continuity independently of facial resemblance. Look for altered hands, jewelry, collars, reflections and background people. Compare the timing of major gestures with the source. Finally listen to the full soundtrack and check alignment around visible speech or impacts. A returned audio track is not proof that every visual beat aligns correctly.
Keep the source, reference set, prompt and review notes together. If a rerun fixes the opening but breaks a later cut, your notes prevent you from mistaking a different failure for progress. Start with a short representative segment before spending time on a complete sequence.
MiniMax H3 Recast FAQ
Is Recast the same as reference-to-video?
No. Recast starts with an existing clip to edit its cast. Reference-to-video starts with a scene brief and reference assets for generating a new scene.
Should I change the set and cast at the same time?
For a first test, keep the set and staging fixed. A single change makes it easier to identify whether identity, motion or scene continuity needs another pass.
How do I handle more than one person?
Label each replacement photo and write an explicit role mapping. Check the provider documentation for its current input limits before submitting the edit.
What should I inspect after a face is covered?
Compare the first clear frame after the occlusion with earlier frames. Also inspect the hand, object edge and background around the transition.
Does this guide show a measured Recast result?
No. The preparation checklist and sample brief are editorial recommendations. The endpoint description is sourced to fal; this guide does not claim a completed generation test.
Turn the brief into your next video
Need a fresh scene rather than a cast replacement? Create your H3 video, use the H3 prompt examples to specify action and sound, and browse video examples for shot ideas. Keep the cast mapping in your project notes so future scenes have a consistent character brief.
Nolan Reeve