AIGC Detector

Upload an image and find out whether it was AI-generated. You get a classifier score, a 4×4 region heatmap, and a check of the file's provenance metadata (C2PA, IPTC, EXIF, PNG generator tags).

browser ──► FastAPI web app (dokploy, this repo /app) ──► Modal GPU backend (modal_app.py, T4)
             · provenance metadata                          · Swinv2 classifier, full image
             · rate limit / auth                            · same model over a 4×4 tile grid
             · HEIC/BMP/TIFF → PNG                          · weight-0 forensic diagnostics

Images are processed in memory and never written to disk.

Model and how to read the score

• Classifier: haywoodsloan/ai-image-detector-deploy, Swinv2 at 256 px, Apache-2.0. Its P(artificial) is the only input to the verdict.
• Verdict bands: ≥ 0.99 likely AI · ≤ 0.10 no strong AI signal · otherwise inconclusive. The upper band is high on purpose. In the n=50 pilot (aigc-detect, 2026-09-22) the clean-set optimal threshold was 0.994, and 20% of real photos scored above 0.5.
• Heatmap: every tile is scored independently by the same model. Hot tiles inside a cool image point to local edits.
• Provenance: reported next to the score, never folded into it. A generator tag is strong evidence. Missing metadata is not evidence of anything.
• Diagnostics: FFT high-frequency energy, noise-residual uniformity, and JPEG grid discontinuity. These carry weight 0.0 and are shown for explainability only.

Known limits: recompression (WeChat, WhatsApp) pulls scores down and noise pushes them up. Recent generators (FLUX dev, Midjourney v7, Firefly 4) evade public detectors. A low score does not clear an image.

API

curl -H "Authorization: Bearer $API_KEY" -F [email protected] https://<host>/api/analyze

The response has likelihood, verdict{key,label}, bands, tiles[{row,col,x,y,width,height,score}] with normalized coordinates, indicators[], metadata{declared_ai,camera,findings[]}, file, and elapsed_s. Web callers without a key are rate-limited per IP. OpenAPI docs are at /api/docs.

The Modal backend also serves VeriDoc's remote_http contract at POST /, so VeriDoc can point AIGC_REMOTE_URL at https://chinesemahim--aigc-detector-api.modal.run/ without code changes.

Run

# backend (one-time: modal secret create aigc-detector-key AIGC_KEY=...)
modal deploy modal_app.py        # if behaviour doesn't change: modal app stop aigc-detector -y, then redeploy

# web app
cp .env.example .env             # fill MODAL_API_KEY
python -m venv .venv && .venv/bin/pip install -r requirements-dev.txt
set -a; . ./.env; set +a; .venv/bin/uvicorn app.server:app --reload
.venv/bin/pytest -q

Deploy: build Dockerfile (container port 8000) with the env vars from .env.example. Production runs as a self-managed compose stack at /root/stacks/aigc-detector on the Contabo VPS, routed by Traefik at https://aigc.rmaikr.top. It is not on dokploy, so a push doesn't redeploy. To ship a change: cd /root/stacks/aigc-detector && docker compose up -d --build. The first request after 5 idle minutes cold-starts the GPU (~20–60 s). The page sends /api/warm on load to hide most of that wait.