---
title: AIGC Image Detector (TechJam 2026 Track 5)
emoji: 🔍
colorFrom: red
colorTo: blue
sdk: gradio
sdk_version: 6.26.0
python_version: "3.12"
app_file: app.py
pinned: false
license: mit
short_description: AI-generated image detector with evidence maps
---

# AIGC image detector — TechJam 2026 Track 5

Live demo of the detector from
[github.com/samt41/techjam_aigsee](https://github.com/samt41/techjam_aigsee):
a frozen DINOv2-B/14 backbone read at blocks 8 and 10 by two small transformer heads
whose logits are averaged. Upload, paste, or pick a sample; you get the probability the
image is AI-generated plus a signed evidence map (red = fake evidence, blue = real).

Headline numbers on the full track demonstration set: clean AUC 0.9978, robust mean
over six corruption families 0.9943, track score 0.9960. Full tables, error analysis
and training code are in the repo.

Sample images: AI-generated ones are from the Community Forensics dataset
(Park & Owens, CC-BY-4.0); real photographs are CC0 from Wikimedia Commons. Credits
in `samples/ATTRIBUTION.md`. This Space runs on ZeroGPU.