Queeenstwest/artwork-detector

AI Artwork Detector

A web application that detects AI-generated anime artwork. It combines a DINOv3 semantic branch and an NPR pixel-fingerprint branch, returns a verdict with calibrated confidence, and abstains on low-confidence inputs instead of guessing.

Graduation thesis — Nguyen Van Hung (102220021), supervised by Dr. Truong Ngoc Chau, University of Science and Technology, The University of Danang.

Single-container deploy: FastAPI serves the API and the React SPA; MongoDB Atlas is the database. First request after a cold start downloads the (gated) DINOv3 backbone, so allow ~1 minute to warm up.

Required Space secrets

Set these under Settings → Variables and secrets:

• —MONGO_URI — MongoDB Atlas connection string
• —JWT_SECRET — a long random string
• —ADMIN_ACCOUNTS — JSON, e.g. [{"email":"admin@example.com","password":"<strong>"}]
• —HF_TOKEN — a token with access to facebook/dinov3-vitb16-pretrain-lvd1689m (gated)
• —CORS_ORIGINS — optional; same-origin needs none, e.g. ["https://<user>-artwork-detector.hf.space"]