MiniMax H3 Recast: What It Changes and What It Keeps

MiniMax H3 Recast replaces people in an existing video with people from reference photos. It is not a text-to-video model wearing a face-swap label. The source video supplies the performance, camera movement, edit points, and soundtrack; the reference photos supply the new cast.

fal publishes it as the separate minimax/h3-max/recast endpoint. AI Live Generator does not currently expose Recast or accept a source-video upload. If you need a source clip first, you can create one here with H3 Max, download it, and take it to a Recast-capable workflow.

Recast in one sentence

Use H3 Recast when the scene already works and the cast does not. It can replace the main people while retaining the source sequence, rather than asking you to regenerate the action around a new prompt.

That makes it a practical fit for a finished or nearly finished shot with useful timing: a two-person exchange, a walk through several cuts, a reaction shot, or a piece of blocking that would be tedious to reproduce. If you still want to change the performance, location, camera, and edit, generating a new video is the clearer choice.

What Recast preserves and what it replaces

“Preserve” should not be read as a frame-for-frame identity guarantee. Recasting still has to synthesize a person across changing angles, expressions, motion blur, occlusion, and cuts. A compatible source and focused instructions may help, but they do not guarantee an exact hairstyle, garment, body shape, or facial detail in every frame. Review the complete video, not just a clean thumbnail from the opening shot.

The current input contract

The current official fal schema defines these controls:

The 30-second video limit and the 15-second shot limit are different rules. A 24-second source can be valid when it contains several shorter shots. A 20-second unbroken take does not satisfy the published contract even though its total runtime is under 30 seconds.

How the people are matched

With no prompt, the endpoint maps the main people from left to right. Imagine a source shot with a chef on the left and a customer on the right. The first reference photo replaces the chef; the second replaces the customer. That mapping continues in the shots where those people reappear.

Left-to-right order can become ambiguous after a cut, a crossing movement, or a change from wide shot to close-up. Use the optional prompt to make the roles explicit:

Replace the chef in the rust-colored apron with Reference 1.
Replace the seated customer in the navy jacket with Reference 2.
Keep the source clothing, choreography, camera, cuts, and sound.

Role descriptions are more useful than “person one” and “person two.” Name something the source video actually shows: position, wardrobe, job, or action. Keep the prompt focused on casting and the details that should remain or change. It is not the place to ask for a different lens, new dialogue, or a rewritten scene.

When Recast is the right tool

Choose Recast when:

• the source performance and edit already communicate the idea;
• you need one to four people replaced across the existing shots;
• keeping the original motion and sound matters more than inventing a new scene;
• you have a clear reference image for every new person.

Choose another workflow when:

• the source action, camera, location, or timing needs to change;
• you only have a still image and no source video;
• you want to continue beyond the end of the source rather than replace its cast;
• you need a dependable, frame-exact likeness for regulated or identity-critical work.

For the final case, use a controlled production and review process rather than treating a generative result as identity verification.

A reference photo is not an opening frame

The reference image tells Recast who should appear. It does not define the opening composition the way an image-to-video first frame does. A front-facing portrait can guide identity while the source actor turns, walks away from camera, or appears in a profile close-up.

Choose a sharp, well-lit photo where the face is unobstructed and recognizable. Use one person per image. A dramatic beauty shot with heavy makeup, sunglasses, or colored lighting may be attractive but gives the model less neutral identity information to carry through difficult frames.

Only upload photos and videos you have the right to use. If the replacement depicts a real person, get their permission and label the finished work where the context could otherwise mislead viewers.

Recast, continuation, and standard generation are different jobs

These workflows may use related H3 models, but their inputs carry different meaning. Recast needs a source video. The standard generator on this site accepts text or still images and cannot inspect an uploaded performance.

Continue with the part you actually need

• MiniMax H3 Recast Reference Photos: How to Map 1–4 People covers photo order, role prompts, crossings, cuts, consent, and per-shot mapping review.
• MiniMax H3 Recast API covers server-side keys, public input URLs, queue jobs, webhooks, and result ownership.
• What Happens to Audio in MiniMax H3 Recast explains why the new face still carries the source voice and when to plan a separate dubbing or lip-sync stage.

They answer different jobs. Start with the tutorial if you are preparing a clip, the API guide if you are building an integration, or the audio guide before recasting dialogue.