photo2cartoon

Why we included this project

This project turns portrait photos into cartoon avatars that keep the person recognizable, which matters more than it sounds: the cartoon style stays relatively realistic instead of sliding into exaggerated anime, and a Face ID loss pulls the output back toward the original identity. The implementation builds on the U-GAT-IT unpaired image translation approach, adding hourglass modules and a Soft-AdaLIN normalization scheme on top. Pretrained weights and an ONNX model are included, so you can run the test script and convert a photo without retraining anything. The preprocessing pipeline (face detection, alignment, cropping, and background removal) is cleanly separated and easy to reuse if you want to study or adapt the approach. One thing to plan around: the released model was trained mainly on young Asian female faces, so other demographics will come out weaker unless you retrain on your own data.