AquaForge8 — dual DINOv2 AIGC detector (TechJam 2026, Track 5)

Two fully fine-tuned DINOv2 ViT-L/14 with-registers towers at 224px into one MLP head. P(AI-generated) per image, trained under real-world degradations with lineage-disjoint validation.

Code, usage and reproduction: https://github.com/bersamin12/TechJam2026-aigc-detection-AquaForge8

python predict.py --images /path/to/dir --out predictions.json

The file dual_d24_ep1.pt is a slim inference export (towers in bf16, optimizer state stripped) produced by scripts/export_finetuned.py.

Downloads last month: -; Downloads are not tracked for this model. How to track