Changing your face in the blink of an eye with AI.

中文文档

Gesture viewfinder + real-time AI restyle. One CUDA GPU runs a FLUX.2 [klein] 4B image API; Mac / browser act as the remote: text-to-image, image-to-image, and a two-hand frame for live styled preview.

Frame your face with both hands → the region restyles live (anime, Ghibli, cyberpunk, …). Close the frame to jump to the next style.

▶️ Demo video (X / Twitter)

• Idea: this LinkedIn post — gesture framing + local model restyle inspired this project.
• Model: Black Forest Labs FLUX.2 [klein] 4B
• Hands: MediaPipe Hand Landmarker

• GPU: server.py (FastAPI + diffusers)
• CLI: gen.py (stdlib only)
• Web: web/ + web-serve.py (static page + same-origin proxy)
• Network: Tailscale / LAN / SSH tunnel — no private addresses in the repo

License
Code: MIT
Default model FLUX.2 [klein] 4B: Apache-2.0 (commercial use OK)
If you set BLINKFACE_MODEL to another weight, that model card’s license applies (some FLUX variants are Non-Commercial).
Follow Black Forest Labs / the model card Acceptable Use policy.

Usage boundary
Only process your own image or people who explicitly consented. Do not use for unauthorized face-swap, impersonation, or anything illegal. Style names (Ghibli, Pixar, …) are descriptive prompt words only — no affiliation with rightsholders.

cp .env.example .env

Existing environment variables win over .env. .env is gitignored.

python -m venv .venv
# Windows: .venv\Scripts\activate
source .venv/bin/activate

pip install torch --index-url https://download.pytorch.org/whl/cu124   # match your CUDA
pip install -r requirements-server.txt
# if your pip diffusers build lacks FLUX.2-klein:
pip install git+https://github.com/huggingface/diffusers.git

# optional pre-download
python dl.py

python server.py
# ready. then:
curl http://127.0.0.1:8000/

On Windows, only open the port you need (e.g. 8000) in the firewall — no need to disable it.

For cross-device access:

# .env
BLINKFACE_BIND=0.0.0.0
BLINKFACE_TOKEN=replace-with-a-long-random-string

• Do not DNAT / expose the GPU port to the public internet.
• Default bind is localhost; without BLINKFACE_TOKEN, do not set BLINKFACE_BIND=0.0.0.0 on untrusted networks.
• web-serve.py injects the token upstream: BLINKFACE_WEB_BIND=0.0.0.0 exposes an unauthenticated generate entry to anyone who can reach that port.
• Details: SECURITY.md.

cp .env.example .env
# BLINKFACE_HOST=http://<gpu-reachable-host>:8000
# BLINKFACE_TOKEN=…   # if the server requires it

curl "$BLINKFACE_HOST/" 

python3 gen.py "a cat holding a sign that says hello world" cat.jpg
python3 gen.py "Transform this person into a Japanese anime character..." out.jpg --img face.jpg

More style prompts: prompts.md.

python3 web-serve.py        # default 127.0.0.1:8080, proxies to BLINKFACE_HOST
open http://127.0.0.1:8080

• Browser only hits same-origin /generate and /health; the proxy injects the token (browser never holds the secret).
• Remote camera: put an HTTPS reverse proxy in front (e.g. tailscale serve --bg 8080).
• To let others on the LAN open the page: BLINKFACE_WEB_BIND=0.0.0.0 python3 web-serve.py (see Security — this exposes the free generate entry).
• Frontend loads scripts/models from jsDelivr / Google Fonts / MediaPipe CDN by default; fully offline needs your own mirrors.

• 1024px / 4 steps ≈ ~1s
• Web defaults to 672×384, 2 steps — tuned for live preview

• Code: MIT
• Default model weights (FLUX.2 [klein] 4B): Apache-2.0
• Other models via BLINKFACE_MODEL: see that model’s card (separate from this repo’s code license)