LandmarkDiff

Anatomically-conditioned facial surgery outcome prediction from standard clinical photography. Upload a face photo, select a procedure, adjust intensity, and see the predicted result in real time.

Powered by MediaPipe 478-point face mesh, thin-plate spline warping, and procedure-specific anatomical displacement models. Runs entirely on CPU. This 2D demo is the foundation -- 3D face reconstruction from phone video is on the roadmap.

Procedure

Select a surgical procedure to simulate

**Rhinoplasty**: Nose reshaping (bridge, tip, alar width) | **Blepharoplasty**: Eyelid surgery (lid position, canthal tilt) | **Rhytidectomy**: Facelift (midface, jawline tightening) | **Orthognathic**: Jaw surgery (maxilla/mandible repositioning) | **Brow Lift**: Brow elevation, forehead ptosis reduction | **Mentoplasty**: Chin surgery (projection, vertical height)
0 100
Example faces (click to load)
  • Front-facing: Use a straight-on frontal photo, not a side profile
  • Good lighting: Even, natural lighting works best. Avoid harsh shadows
  • Neutral expression: Keep a relaxed, neutral face for accurate landmark detection
  • No obstructions: Remove glasses, hats, or anything covering the face
  • Resolution: At least 256x256 pixels. The image will be resized to 512x512 internally
  • Single face: Make sure only one face is clearly visible in the frame
LandmarkDiff v0.2.2 · TPS on CPU · MediaPipe 478-point mesh · GitHub · MIT License · For research and educational purposes only