MiniMax H3 Recast: New Faces, Same Take, Step by Step
Swap the people in a clip you already shot and keep the take. Settings, photo picks, multi-person ordering and copy-ready prompts for MiniMax H3 Recast on each::labs.
By Eftal Yurtseven · Nov 17, 2026 · 9 min read

You already have the take. The dolly move landed, the actor hit the mark, the room tone is clean, the cut on the turn feels right. Then someone asks for the same spot with a different face in it. Reshooting means a crew, a location and a day you don't have. Regenerating from scratch means losing every one of those small things that made the take worth keeping.
That's the job MiniMax H3 Recast is built for. MiniMax H3 Max Recast takes an existing clip and one photo per new person, and hands the clip back with those people swapped in while the motion, camera work, cuts and audio stay where they were. You aren't asking for a new video. You're asking for the same video, recast.
This guide walks through it step by step on each::labs: what the source clip has to look like, how to pick the photos, how the default left to right ordering works, how to write the optional prompt when the default isn't what you want, and when to pick 768P over 1080P.
What MiniMax H3 Recast Actually Does
Recast sits in the MiniMax H3 family, the multimodal model MiniMax open-sourced in August 2026. MiniMax describes H3's editing as the ability to replace, add or remove people and objects while untouched content stays stable. Recast narrows that down to one very specific promise: replace the people, keep the take.
The inputs are short. A video_url for the source clip. A reference_image_urls list with one to four photos, one per new person. An optional prompt of up to 2,000 characters. A resolution of 768P or 1080P, with 1080P as the default. And an optional seed. The output is an MP4 with the new cast in place.
Here's the part worth sitting with. Most video models start from nothing and try to reach your idea. Recast starts from your idea, already shot, and changes as little as possible. That's a different kind of control, and it's why the clip you feed it matters more than any prompt you write.

Step 1: Prepare a Source Clip That Recast Will Accept
The rule from the schema is precise: the source video runs 5 to 30 seconds, and no single shot inside it can be longer than 15 seconds. So a 24 second clip made of two 12 second shots is fine. A 20 second clip that holds one unbroken shot is not, even though the total length fits.
That shot rule is the one people trip over. If your best footage is a long oner, cut it before you upload, or trim it down to 15 seconds. If you're working with an edited sequence, check the longest shot, not the running time.
Pick footage where the people you want to replace are clearly visible for most of the clip. The model page itself advises checking identity through cuts and occlusions, which tells you where the hard moments are: someone walking behind a pillar, a hand crossing the face, a cut to a wide where the face is tiny. You don't have to avoid those moments. You do want to know where they are so you watch them first when the result comes back.
Host the file at a direct, retrievable URL and keep it available while the job runs. A link that expires halfway through processing is the most boring way to lose a run.

Step 2: Choose Reference Photos That Carry One Identity Each
Each photo in reference_image_urls stands for one new person. One photo, one identity. Don't send two angles of the same person as two entries; Recast will read them as two different people and try to place both.
The each::labs model page asks for reference images with clear subjects and compatible framing, and that's the whole brief in six words. A face that fills a good part of the frame, lit evenly, not hidden by sunglasses or a hand, reads far better than a group shot cropped down to one blurry head. Compatible framing means the photo shouldn't fight the clip: if the person in your footage is seen from the chest up, a head-and-shoulders portrait is a better match than a full-body shot from across a street.
A strong opinion here: one great photo beats a creative one. The reference isn't the place for mood lighting or a dramatic angle. The clip already supplies the lighting, the angle and the drama. All the photo has to do is say who.
Step 3: Get the Multi-Person Order Right
With more than one person in frame, ordering decides everything. The default behavior, straight from the schema, is that the photos replace the main people in the video from left to right, and each new person replaces their counterpart in every shot they appear in.
So if your clip shows two people at a table, the first photo in the list replaces whoever sits on the left, the second replaces whoever sits on the right. If the clip cuts to a close-up of the person on the right, the second identity follows them into that shot too. You don't have to set anything for that. Order the list to match the frame and the default does the work.
The default breaks down when left and right stop meaning one thing. People cross the frame. A reverse angle flips who's on which side. A third person walks in halfway through. That's when you stop trusting position and start using the prompt.

Step 4: Write the MiniMax H3 Recast Prompt for Who Becomes Whom
The prompt is optional, and the schema says exactly what it's for: who becomes whom, or what else to keep or change. Treat it as a casting note, not a scene description. The scene already exists. Describing it again only gives the model room to drift.
The simplest useful prompt maps photos to people by something you can see in the clip. Clothing and position work best because they don't change between shots.
The woman in the red coat becomes the person in photo 1. The man in the grey suit becomes the person in photo 2. Keep everything else exactly as it is.
When people move, anchor them to what they do rather than where they stand:
Photo 1 replaces the person who opens the door at the start. Photo 2 replaces the person sitting on the sofa who stands up to greet them. Each keeps their own role in every shot, including the reverse angle.
When you want to recast only part of the scene, say who stays:
Replace only the barista behind the counter with the person in photo 1. The customer at the register stays as the original person. Keep the camera move, the background and the sound unchanged.
And when you want something beyond the face carried over, name it, and name what must not change:
The lead runner becomes the person in photo 1 and keeps the same running gear as in the original. The runners behind stay as they are. Do not change the track, the crowd, the camera tracking or the audio.
A few habits keep these prompts sharp. Refer to photos by number, in the order you put them in the list. Describe people by what is visible and stable. Say "keep" out loud for the things you care about most, even though keeping is the default. You have 2,000 characters, but the good prompts here rarely need more than three sentences. Length invites interpretation, and interpretation is the thing you came to Recast to avoid.
Step 5: Pick 768P or 1080P
The resolution field takes two values, 768P and 1080P, and 1080P is the default if you send nothing.
Use 768P when you're testing. The first runs on a new clip are about checking the casting, the ordering and the hard moments, and you don't need full resolution to see whether the right person ended up on the left. Use 1080P for the version you'll actually ship, cut into a timeline or show a client.
That split also lines up with the seed. Leave seed empty and you get a random one each time. Once a test run looks right, note the seed, keep the prompt and photos identical, and rerun at 1080P. Changing one setting at a time (the model page's own advice) is how you find out what actually moved the result.

Step 6: Run It on each::labs
You can run Recast from the model page on each::labs, or call it through the same API as every other model in the catalog. A request looks like this:
curl -X POST https://api.eachlabs.ai/v1/prediction \
-H "Authorization: Bearer $EACHLABS_API_KEY" \
-H "Content-Type: application/json" \
--data '{
"model": "minimax-h3-max-recast",
"input": {
"video_url": "https://example.com/cafe-scene.mp4",
"reference_image_urls": [
"https://example.com/new-barista.jpg",
"https://example.com/new-customer.jpg"
],
"prompt": "The barista behind the counter becomes the person in photo 1. The customer at the register becomes the person in photo 2. Keep the camera move, the cafe and the sound unchanged.",
"resolution": "768P"
}
}'The response is a prediction ID, not a video. Poll /v1/prediction/{id} for the result, or pass a webhook_url and let each::labs call you when it's done. Build for that from the start, especially if you plan to recast a batch of clips rather than one.
If Recast is one stage of a larger pipeline (generate a base clip, recast it for three markets, reframe it for vertical) it fits naturally into an each::labs flow, where each step hands its output to the next.
Step 7: Review the Whole Clip, Not the Thumbnail
A good thumbnail proves very little. The real check happens across the full running time, and the model page lists the places to look: identity through cuts and occlusions, motion continuity, camera changes, the ending and the audio.
Watch the first frame after every cut, because that's where a new identity has to be picked up again. Watch the moment someone turns away from camera and back. Watch the last second, since endings are easy to skip. Then listen once with your eyes closed. Recast is meant to keep the original audio, and the take you started with should still sound like that take.
If one person is wrong, fix the ordering or tighten the prompt for that person only. If both are wrong, check the photos first. More often than not the fix is a better reference, not a longer prompt.
Where Recast Fits in the MiniMax H3 Lineup
Recast is for when the footage exists. If it doesn't yet, start somewhere else. MiniMax H3 Max Reference to Video builds new clips around reference images, and our step by step guide to H3 Reference to Video covers how to keep one character steady across shots. For the bigger picture of what H3 accepts and why references matter so much in this family, read MiniMax H3 Speaks Your Visual Language.
The two approaches pair well. Generate a scene with a placeholder cast, get the blocking and camera right, then recast it for each version you need. One approved take, many faces. If you want to compare Recast with other tools that change what's already in a clip, the video-to-video models on each::labs sit together on one page, and the full model catalog is a click away. MiniMax's own H3 announcement is worth a read for how the family came together.
The take was always the expensive part. Recast lets you keep it.
Frequently Asked Questions
What is MiniMax H3 Recast used for?
Replacing the people in footage you already have. You give it a source clip of 5 to 30 seconds and one photo per new person, and it returns the clip with those people swapped in while the motion, camera work, cuts and audio stay the same. It's the tool for making versions of an approved spot, trying a recast before you produce more assets, or personalizing an existing clip.
How many people can MiniMax H3 Max Recast replace at once?
Up to four. The reference_image_urls field takes one to four photos, one per new person. By default the photos replace the main people in the video from left to right, and each new person follows their counterpart into every shot they appear in. Use the prompt when that left to right order doesn't match who should become whom.
Why does Recast reject my clip?
Usually because of the shot rule rather than the total length. The source can run 5 to 30 seconds, but no single shot can be longer than 15 seconds. A clip that holds one unbroken shot for 20 seconds fails even though the total fits. Cut or trim the long shot and upload again.
Do I need to write a prompt at all?
Not always. With one person in frame, or several people who stay on their side of the screen, the default left to right mapping often does the job. Write the prompt when people cross the frame, when only some of them should change, or when you want to say plainly what else to keep. Keep it short and refer to photos by number.



