Fleet: Few Shots Lead Effective AI-generated Image Detection

Accepted by ICML 26!

Paper on Arxiv

https://arxiv.org/abs/2606.31082

Treasure dataset

Apply for access on huggingface: https://huggingface.co/datasets/ThreeLiu/Treasure

Fleet code

How to run?

1. Open Run.sh and fill in the paths under the "Required path configuration" section at the top of the file (DINOv3 weights, AIGIBench train/val sets, few-shot fake/real data directories, etc.).
2. Make sure PYTHON points to your Python interpreter and ALL_GPUS lists the GPUs to use.
3. Run:

bash Run.sh

Run.sh runs three stages in sequence: Stage 1 pretraining -> Stage 2 few-shot fine-tuning -> Stage 3 summary. Logs, checkpoints, and the summary are written under OUTPUT_ROOT (configured in Run.sh).

The pretrained and few-shot model weights are available at: https://pan.baidu.com/s/1o07s69fOXVWUcJLiO3YftA?pwd=b6un and https://huggingface.co/wangjiaan/Fleet .