Product guide · Updated 20 October 2026

Four people on screen.
Replace only your two.

Assign different photos to roles A and B, and leave everyone else unchanged. SwapWeave uses original-person identity references to decide who gets replaced, with an independent track for each visible face.

Original: four people together — Original and engine output at the same source timestamp. Only the leftmost woman and rightmost man have replacement bindings; the two people in the middle remain unchanged.

Result: the leftmost woman and rightmost man are replaced, while the middle two people remain unchanged — Original and engine output at the same source timestamp. Only the leftmost woman and rightmost man have replacement bindings; the two people in the middle remain unchanged.

Choose your own on-screen pairing.

Bring two familiar faces into the same story by assigning a different photo to each selected role. The identity reference comes from the original person in the video. The replacement photo supplies the new appearance; it does not decide who the original person is.

How do you bind multiple people?

Pause on a clear frame and choose Pick face.
Select the original face for role A and assign replacement photo 0.
Select the original face for role B and assign replacement photo 1.
Finish picking and continue playback. Unbound people remain unchanged.

Each role has its own enabled state, recognition threshold and original-face reference set. Disabling A does not change B's photo. Selecting another video window preserves saved roles and clears temporary tracks so identities can be confirmed again.

What if people change places?

Assignments do not follow left-to-right order. The engine compares each original video face with role references and uses competing-role separation, tracking and consecutive observations to confirm identity. A change in detector output order does not directly exchange A and B.

Head rotations, occlusion or cuts may still leave too little evidence. The matching matrix shows each face's actual similarity to A and B, the role thresholds and the best reference sample. This distinguishes a missed detection from a detected face that was not recognized.

Can the same person appear twice?

A person holding their own photograph, or appearing more than once, creates separately tracked face instances. Each instance is confirmed independently. Both can then belong to the same role and use the same replacement. Confirmation from one instance does not automatically authorize a newly appearing instance.

Different views of one person can be added to the same role, up to 12 original-face references. They share a replacement photo and threshold, and can be deleted individually. Reference count, role count and simultaneous replaced-face count are separate concepts; the current limit is four replaced faces per frame.

Try the role-mapping logic

ROLE MAPPING / 04 SUBJECTS↗

A → 01

B → 02

C + D remain unchanged

Interactive mapping illustration, not rendered face-swap footage. Try the toggles or switch positions.

Which multi-person performance numbers matter?

Measure detection, identity recognition, and generation plus blending for all replaced faces, then inspect delivered output rate and capture-to-display delay. A two-face workload does not establish four-face performance. A cheap frame with no replacement is not a successful face-swap result.

The prototype has four-face still-fixture and selected moving-clip tests. Complex multi-person footage still needs evaluation by device and clip. Read the role-binding explanation and measurements with their limits.