AI workflows
Map both identities in a multiple face swap video
A multiple face swap video needs two separate identity assignments, not merely detection of several faces. Map each source person to a target performer, mark where faces cross or disappear, then inspect the first returning frames. Reject an identity exchange even when both individual faces look convincing.
In this guide
- OpeningConfirm each target performer is associated with the intended source identity.
- CrossingTrack the people as their screen positions and relative sizes change.
- ConcealmentMark what is hidden instead of claiming evidence from an invisible face.
- ReappearanceCompare the first clear returning face and the first frame after each cut.
Identify people independently of screen position
Start with names for the two source identities and separate labels for the target performers in the footage. For this example, source Mira maps to the performer carrying a bag, while source Theo maps to the performer waiting near the window. The clothing and prop are tracking cues in this clip; they are not proof of facial identity.
Record a clear reference moment for each target. Choose a point where that person's face can be inspected and distinguish it from the other face in the frame. Use the actual clip timestamp or frame number in your working notes. Do not fabricate universal reference positions or assume the first frame is always suitable.
Screen-left and screen-right are temporary observations. The bag carrier may cross the room and end on the opposite side. A map that says only “Mira equals left” becomes wrong as soon as positions change. Track the performer through the event, then compare the assigned source identity when the face becomes visible again.
| Source identity | Target performer | Reference moment to fill in | Visible tracking cue |
|---|---|---|---|
| Mira | Person entering with the bag | Clear face before crossing: ____ | Bag and entrance action |
| Theo | Person waiting by the window | Clear face before response: ____ | Window position and waiting role |
Check selection separately from identity assignment
FaceFusion's selector documentation exposes one, many and reference modes, ordering controls, a reference frame and a reference-face position. These are specific implementation controls. They provide evidence about selecting detected faces, not proof that a many-face setting creates two different source-to-target identity mappings.
Before a trial, ask the selected tool's documentation a narrow question: how are two distinct source identities assigned to two distinct target performers? If it only explains replacing detected faces with one source, that does not establish the operation your map requires. Do not infer the answer from a plural label or a demo with several people.
Face ordering can help describe which detected face is selected in a particular frame. It is not a permanent identity label through a crossing. The largest or leftmost face may belong to a different performer later. Keep the original source-to-target map as the review authority, even when a product exposes ordering controls.
Write a visibility log before generating the edit
Watch the original footage and mark each person's face as clear, partly covered, hidden or returning. Use intervals where useful rather than collecting arbitrary screenshots. In the station example, the bag carrier passes in front of the waiting person, briefly covering the waiting person's face. The first clear frame afterward is a priority review point.
Also mark head turns. A profile or back-of-head moment changes the evidence available even without another person crossing. Do not score a fully hidden face as preserved merely because the torso remains recognizable. The visible body can help track the performer, but it cannot reveal the hidden facial result.
Include cuts in the same log. A cut can change angle, relative size and screen position at once. Treat the first clear face after the cut as a fresh assignment check. A mapping that survives an uninterrupted approach may still fail when the new shot shows the people in a different arrangement.
| Event in source footage | Mira target visibility | Theo target visibility | Review priority |
|---|---|---|---|
| Before the crossing | Clear | Clear | Confirm both assignments |
| Bag carrier passes the window | Clear or partly turned | Partly covered or hidden | Trace each performer’s path |
| First moment after crossing | Clear | Returning | Compare the returning identity |
| New camera shot | Reassess | Reassess | Check both assignments again |
Use mask documentation for region questions
FaceFusion documents occlusion and region masks separately from face selection. That separation is useful: choosing a target face and deciding which visible region to process are different operations. Mask controls do not by themselves establish which source identity belongs to which person.
For a hand passing over a cheek, the review question is whether the hand and face remain physically readable in the result. For a face disappearing behind a shoulder, the question includes which identity reappears. A mask might be relevant to the visible boundary, while the identity map remains relevant to assignment.
Do not present an occlusion control as automatic recovery of hidden facial evidence. Where the face is fully covered, the source frame does not show its features. The strongest available check may be the last clear frame before concealment and the first clear frame afterward, together with the performer's continuous body path.
Choose a test that includes the risky event
A trial containing only the opening faces cannot answer whether the mapping survives the crossing. Select a short permissioned segment that includes a clear view before the event, the event itself and a clear reappearance. Keep the scope narrow enough that you can inspect every transition that matters.
The proposed brief below is an assignment and review plan, not a claim that a particular product supports the full replacement. Confirm that capability from the current selected operation first. If the tool cannot document separate mappings, stop at qualification rather than inventing steps for a simultaneous two-person edit.
Proposed two-identity test brief
Map source Mira to the bag-carrying performer and source Theo to the performer waiting beside the window. Include the crossing and the first clear view of both faces afterward. Compare each returning face with its own source. Keep clothing and props as tracking cues, and reject a result where the identities exchange.
Review performance first, then each identity
Watch the modified segment at normal speed to understand who moves where. Then follow one performer at a time through the same segment. Compare that performer's clear face before the crossing, last visible face before concealment and first visible face afterward. Repeat for the other performer; do not let a convincing nearer face distract from the farther one.
Write defects with an event and an identity: “Theo target returns with Mira's likeness after the shoulder overlap.” That is more actionable than “faces look strange.” Also record an unrecognized face separately from an exchanged face. Both can fail the shot, but they suggest different qualification or revision questions.
Review the final frame and any cut again after the individual passes. Check the bodies, hands and props as well. Correct identity assignment does not make a broken embrace acceptable, just as plausible anatomy does not excuse exchanging the people.
- Confirm both source-to-target assignments at the opening.
- Follow each performer through the crossing independently.
- Inspect the last clear and first returning face around concealment.
- Recheck both people after every cut and at the ending.
- Reject swapped identities even if each face looks individually plausible.
Keep product names and edit scopes distinct
Higgsfield's November 2025 swap guide warns about nearest-face targeting in its older workflow. Use that as a reason to verify current targeting, not as a universal statement about every face editor. A newer operation may have different controls and still needs evidence for the task you intend.
Viggle Multi terminology describes a whole-character workflow; it is not evidence of multi-face swapping. Likewise, a face editor's many mode is not automatically a two-source assignment system. Qualify the actual operation instead of combining labels from different products into an unsupported tutorial.
Zombie Couple currently offers a local two-photo casting preview and reference-film playback; generation and checkout are disabled. This multiple face swap video guide supplies an original mapping and audit method from documented selection concepts. No paid render, simultaneous two-person replacement or automated concealment repair was tested.