FiSeR: Fine-Grained Source Representations for Cross-Domain AI Image Detection

Introduction

FiSeR learns transferable representations for AI-generated image detection with two supervised contrastive objectives: a coarse natural/synthetic objective and a fine-grained objective over synthetic source identities. The released encoder is based on DINOv3 ViT-L/16 and supports GPU-FAISS kNN, prototype, linear, and SVM heads.

This repository contains the FiSeR training and evaluation code, single-image inference, and source snapshots with common launchers for 16 comparison methods. Baseline weights, datasets, embeddings, and experiment outputs are not included.

Dataset

FiSeR is trained on WildFake and evaluated on WildFake, Community Forensics, AIGIBench, Chameleon, and GenImage. Obtain each dataset from its official release and follow its license. The expected LMDB format and split sizes are documented in DATASETS.md.

The repository provides adapters for all five datasets; no prepared LMDB or image is redistributed. The full Community Forensics training release contains 4,783 source categories and can be processed directly from its parquet shards:

hf download OwensLab/CommunityForensics --repo-type dataset \
  --local-dir /path/to/CommunityForensics

python scripts/prepare_lmdb.py \
  --dataset community --root /path/to/CommunityForensics \
  --split train --output /path/to/community_train.lmdb \
  --expected-count 2971612 --expected-source-count 4783

See DATASETS.md for the required raw layouts and exact commands for WildFake, Community Forensics, AIGIBench, Chameleon, and GenImage.

Installation

Python 3.10+, CUDA, and a GPU-enabled FAISS build are required for formal evaluation and kNN inference.

conda create -n fiser python=3.10 -y
conda activate fiser
pip install -r requirements.txt
pip install -e . --no-deps

Verify that FAISS sees at least one GPU:

python -c "import faiss; assert hasattr(faiss, 'GpuIndexFlatIP') and faiss.get_num_gpus() > 0"

DINOv3 requires transformers>=4.56.1. Depending on the current access policy, the base model may also require hf auth login.

Inference

Single-image inference runs the encoder and classification head on CUDA. Load a local FiSeR checkpoint together with a trained linear head as follows:

CUDA_VISIBLE_DEVICES=0 python infer.py \
  --image /path/to/image.jpg \
  --model-name facebook/dinov3-vitl16-pretrain-lvd1689m \
  --checkpoint /path/to/fiser_checkpoint.pth \
  --head linear \
  --head-checkpoint /path/to/linear_head.pt \
  --layer 18 \
  --device cuda

For the published encoder and a GPU-FAISS kNN feature bank:

CUDA_VISIBLE_DEVICES=0 python infer.py \
  --image /path/to/image.jpg \
  --model-name heyongxin233/FiSeR-DINOv3-ViT-L16 \
  --head knn \
  --database-pt /path/to/wildfake_train_embeddings.pt \
  --layer 18 --k 25 --faiss-devices 0 --device cuda

A portable linear or prototype head can be built from an existing embedding archive with scripts/build_inference_head.py. The inference command prints a JSON prediction using the label convention natural=0, synthetic=1.

Testing

First extract normalized hidden-state embeddings:

CUDA_VISIBLE_DEVICES=0 python extract_embeddings.py \
  --model-name heyongxin233/FiSeR-DINOv3-ViT-L16 \
  --dataset-lmdb /path/to/test_lmdb \
  --output /path/to/test_embeddings.pt \
  --layers 15 16 17 18 19 20 21 22 23

Then evaluate a training feature bank against a test archive. evaluate.py rejects CPU-only FAISS installations.

CUDA_VISIBLE_DEVICES=0 python evaluate.py \
  --database-pt /path/to/wildfake_train_embeddings.pt \
  --test-pt /path/to/test_embeddings.pt \
  --head knn --layers 18 --max-k 25 --faiss-devices 0 \
  --output /path/to/metrics.json

Run the unit tests with:

pytest -q

Training

Update train_lmdb in configs/fiser_wildfake.yaml, then launch distributed training:

CUDA_VISIBLE_DEVICES=0,1 torchrun --nproc_per_node=2 \
  train.py --config configs/fiser_wildfake.yaml

The effective global batch size must be divisible by the distributed per-step batch size. The training entrypoint saves standard Hugging Face checkpoints, which can be loaded directly by the extraction and inference commands.

Baselines

The baseline release is source-only: no trained baseline weights are distributed. The baselines/ directory contains common GPU launchers for all 16 comparison methods. The public training example defaults to five epochs; change the environment variables as needed for your own experiment.

TRAIN_LMDB=/path/to/train.lmdb \
EPOCHS=5 \
OUTPUT_DIR=/path/to/output \
CUDA_VISIBLE_DEVICES=0 \
baselines/scripts/train.sh resnet50

LMDB_PATH=/path/to/test.lmdb \
MODEL_PATH=/path/to/your_checkpoint.pth \
CUDA_VISIBLE_DEVICES=0 \
baselines/scripts/evaluate.sh resnet50

Use the method ID in baselines/README.md to train and test CNNDetection, LGrad, Gram-Net, FreqNet, NPR, SAFE, LASTED, UniFD, AIDE, DIRE, Effort, SPAI, C2P-CLIP, or LOTA in the same way. The source snapshots under baselines/vendor/ retain their native dependencies and licenses.

Citation

@article{zhang2026fiser,
  title={FiSeR: Fine-Grained Source Representations for Cross-Domain AI Image Detection},
  author={Zhang, Shan and He, Yongxin and Zhang, Mingming and Tian, Huiwen and Ma, Lei},
  journal={arXiv preprint arXiv:2606.00606},
  year={2026}
}