FaceApp-style gender swap for photos. FaceApp's own model is proprietary, so this uses the open-source approach that gives the same kind of result: a StyleGAN2 face generator edited in latent space.
- Detect & align the face with OpenCV YuNet, cropping it the way FFHQ (StyleGAN2's training set) was cropped.
- Invert the crop into StyleGAN2's W+ latent space with the e4e encoder.
- Pivotal Tuning (PTI): briefly fine-tune the generator on your face so the result still looks like you (skipped in
fastmode). - Edit: move the latent along a learned gender direction (source).
- Paste back the swapped face into the original photo with a feathered blend.
Requires Python 3.9 - 3.12; an NVIDIA GPU is strongly recommended (4 GB is enough).
Just run the launcher for your shell. On first run it creates the psdenv virtual environment,
installs requirements.txt (again whenever that file changes) and downloads any missing models
(about 1.3 GB, into pretrained/). Arguments are passed through to main.py.
run.bat # Windows (double-click or cmd)
.\run.ps1 # PowerShell
./run.sh # Linux / macOS / Git Bash
run.bat photo.jpg --quality best
If PowerShell refuses to run scripts: powershell -ExecutionPolicy Bypass -File run.ps1.
To only fetch the models: python main.py --download-models.
python main.py # web UI
python main.py photo.jpg # writes photo_swapped.jpg
python main.py photo.jpg --to female --strength 1.3 --quality best -o out.png
| Option | Meaning |
|---|---|
--to auto|female|male |
Target gender. auto (the default) flips whatever is detected. |
--strength |
How far to push the edit (1.0 = full swap). |
--quality fast|balanced|best |
fast = encoder only (a few seconds); balanced/best fine-tune on your face for 100/250 steps (~1 / ~3 min on a GTX 1050 Ti). |
--save-crops |
Also save the aligned input, reconstruction and edit side by side. |
In the web UI you don't have to pick a direction: when you upload a photo, a gender classifier (FairFace) checks the face and selects the opposite, e.g. "Detected male (99%), so to female is selected". You can still change it to To female, To male or Auto before pressing Swap.
The code here is MIT-derived (e4e, rosinality/stylegan2-pytorch). The pretrained StyleGAN2 FFHQ weights are under NVIDIA's non-commercial license, so this is for personal/research use.
