Deepfake Detector

Upload an image or short video to check if it's AI-generated. Powered by a Swin Transformer V2 model fine-tuned on real vs synthetic images.

How it works

• Images: a single forward pass through microsoft/swinv2-base-patch4-window16-256 with a 2-class fine-tuned head (label 0 = real, label 1 = fake). The reliability score is 1 - P(fake).
• Videos: 5 frames are sampled uniformly across the duration; the per-frame P(fake) values are averaged.

Endpoints

• GET / — the web UI (English / French toggle)
• POST /api/predict — multipart upload, field file. Returns { media_type, p_fake, reliability, n_frames, frame_probs? }.

Local development

docker build -t openfake-demo .
docker run --rm -p 7860:7860 openfake-demo

Then open http://localhost:7860.

Notes

• Free Hugging Face Spaces sleep after 48 hours of inactivity. The first request after sleep takes 30–60 seconds while the container boots.
• Uploaded files are processed in memory and not persisted.