Product guide · Updated 20 October 2026
Real-time face swap.
In your Mac video window.
SwapWeave is a native Mac face-swap prototype. Select a playing video window, assign your chosen photos to specific roles, and manage the cast from a floating toolbar. Frames and references are processed on your device.
Recast a story you already love.
Try a different casting choice on a personal rewatch: assign a favorite photo to role A, then another photo to role B. Face replacement changes their on-screen appearance while the source player continues the audio, story and camera work.
Use the player you already have. After you select its window through macOS screen sharing, SwapWeave displays the processed frame at that window's position. Mouse input passes through to the original player.
How to get started
- Select the window.Choose Window in the floating toolbar, grant the system screen-capture permission and select the video window.
- Bind the original person.Choose Pick face, click a detected face in the original frozen frame, then select a clear replacement photo.
- Add the other roles.Give different people different photos. Faces you have not bound remain unchanged.
- Press play.Finish picking and continue watching. Collapse the toolbar to a circle; role references, replacement photos and thresholds are saved automatically.
Watch the 28-second walkthrough ↗
Why might the original face briefly return?
Detection finds the face; recognition confirms that it belongs to a bound role. Blur, occlusion and different head angles can affect either stage. Matching diagnostics show the actual similarity between each original video face and each role, along with acceptance and rejection reasons.
Add original-face references of the same person at other angles, then use measured scores to adjust that role's recognition threshold. Detection and recognition controls span 0.10–0.95. Lower values admit more candidates, including potential mistakes. A role can hold up to 12 original-face references, each available for inspection and deletion.
Mac implementation and validation scope
| Interface and core | Native macOS overlay; C++ / Objective-C++ / Metal |
|---|---|
| Inference and composition | Core ML inference, Metal image processing and blending |
| Data | Original frames, role references and replacement photos stay on-device |
| Multiple people | Up to four replaced faces per frame; separate photos for separate roles |
| Real-time target | 1080p, 30 fps, 2–4 people; sustained end-to-end validation on complex moving video is ongoing |
A four-face repeated-still fixture measured 18.19 ms mean and 24.09 ms P95 engine processing. Those figures do not equal delivered screen-playback fps or total capture-to-display delay. Read the complete measurement scope.
What should an evaluation check?
Minimized windows, players that stop background rendering and protected video may not be capturable. Large head rotations, occlusion and identity continuity need evaluation on the actual footage. The current prototype is available for partnership discussions; commercial distribution needs appropriate model rights and an agreed integration scope.