Face Replace in a Video With Several People: Who Gets Replaced, and How to Choose (October 2026)
Most video-edit models replace every face with one avatar. Recast maps one photo to each person, left to right, and Kling O3 lets a prompt name the target with @Element1. How each handles two people, what to write in the prompt, and what to do when a model gets it wrong.
Replace-the-person tools usually assume there is one person. Paste a Reel with two people talking, or a couple dancing, and a model that replaces "the person" has to decide which, or all. How a tool decides determines whether the result is a clip you can post or a clip where both people have become you. This guide explains the three behaviours you will meet, how to steer each, and the checks that catch a wrong choice early.
Three behaviours
1. Every face becomes one avatar. This is how HappyHorse video-edit works, the model our autopilot runs today. You supply one identity and every face it finds is replaced with it. For a clip with one subject this is right; for two people it produces twins. That is why the Studio warns before running when it detects more than one face, and why we describe the case honestly rather than hiding it (HappyHorse video-edit).
2. The prompt names the target. Kling's O3 editor lets you define characters as elements and refer to them in the prompt, with the source as @Video1 and the replacement as @Element1. A prompt such as "Replace the woman on the left in @Video1 with @Element1; leave the man unchanged" is a legitimate instruction. How reliably the model obeys an instruction about position and which person, we have not yet measured (Kling O3 edit).
3. One photo per person. MiniMax H3 Max Recast takes one to four reference photos and replaces people left to right: the first photo replaces the leftmost person, the second the next, and so on. The order can be overridden in the prompt: "use the first reference photo for the presenter on the right". For a two-person clip where you want to replace only one person, you supply one photo and, if needed, a sentence that says which. It is the first model we have seen where "who becomes whom" is an explicit input rather than a hope (Recast).
Which to use
| Situation | Best fit |
|---|---|
| One person in the source | any; the cheapest that holds your face |
| Two people, replace one | Recast with one photo and a position in the prompt, or Kling O3 with a named element |
| Two people, replace both with two different people | Recast with two photos |
| A group or crowd | none reliably; trim to the subjects or accept the original background people unchanged |
| Two people, you want both to be you | a model that replaces every face with one avatar |
How to write the instruction
- Name people by something visible and stable: position (left, right, foreground), clothing colour, role (the host, the guest). Avoid names; the model cannot see names.
- Say what to leave unchanged, not just what to change.
- Use the order the model documents. With Recast, supply the photos in left-to-right order of the people in the frame, or say explicitly which photo goes where.
- Keep to the first seconds for the test. The ambiguity shows in the first shot.
Checking before you pay for the whole clip
A five-second test costs $0.70 to $1.50 across the three models. Watch for three things:
- Did the right person change? Check the first frame and the frame after the first cut.
- Did the other person stay themselves? Sometimes the unchanged person's face drifts too, because the model regenerates more than the target.
- Does identity hold when they cross or overlap? Two people passing each other is where models swap or merge faces.
If the first check fails, do not run longer. Change the instruction, or crop the source so only one person remains.
What we have not verified
We have not yet run the two newer models on a two-person source. The explicit mapping is in the schemas and in the vendor documentation; whether it works reliably on real footage with real faces is the question our planned test (about $1.50 for a 5-second two-person clip) is meant to answer. Treat any claim, including our own description above, as a specification until a paired run says otherwise.
Rights with two people
Every person in a source video is a likeness. Replacing one with you does not release the other; keep the unchanged person's consent in mind, especially if they are a private individual. Using clips with minors is a case where we advise not to proceed at all.
Common questions
What happens if I do nothing about the second person?
With HappyHorse video-edit, every face becomes your avatar. With Recast or Kling you can name or map the target, but you should verify on a short test.
How do I refer to a person in a prompt?
By position, clothing or role: the woman on the left, the man in the blue jacket, the host. Models cannot see names.
What if the two people cross?
That is where models swap or merge faces. Test the moment of overlap before running the full clip.
How much does a test cost?
About $0.70 to $1.50 for five seconds across the three models.
Is it fine to replace one person and leave the other?
Technically possible, but the other person remains a likeness. Keep their consent in mind, especially for private individuals.
How to do this without code: The autopilot warns you before it runs when it detects more than one face in the source. In AutoHustle this is the autopilot: paste the Reel link, add one photo of yourself and confirm the plan (about 4 credits per second of finished video, so 60 credits for a full 15-second call; analysing the Reel is free). See How to put your face in any Reel and Recreate an Instagram Reel.
Related reading: How to Replace a Person in a Video With AI: Three Models, What Each Costs, and the Mistakes That Break It (October 2026) · AI Dance Video From One Photo: Motion Transfer or Recast? A Decision Guide (October 2026)
For developers: the fal endpoints
| Model | fal endpoint |
|---|---|
| MiniMax H3 Max Recast | minimax/h3-max/recast |
| Kling O3 Pro video-to-video edit | fal-ai/kling-video/o3/pro/video-to-video/edit |
| HappyHorse video-edit (1.0) | alibaba/happy-horse/video-edit |
Recast: reference_image_urls (1–4, one per person, replaced left to right). Kling O3: @Video1 and @Element1 in the prompt. HappyHorse: up to five references cited as @Image1…@Image5; all faces replaced with the avatar.