PRISM: PhotoRealistic Image Synthesis and Manipulation Benchmark for Generated Image Detection (CVIU 2026)

Filippo Bartolucci · Samuele Salti · Giuseppe Lisanti

Computer Vision Lab, University of Bologna, Italy

---

Installation

pip install -r requirements.txt

A CUDA-capable GPU is required. Install the appropriate PyTorch build for your CUDA version from pytorch.org before running the above.

---

Dataset

The PRISM dataset is available on HuggingFace.

The training code expects the following folder layout:

<dataset_root>/
  imgs/                        # real images
  imgs_<GeneratorName>/        # generated images for each generator
  imgs_<GeneratorName2>/
  ...

Update the absolute paths at the top of dataset.py to point to your local copy:

PRISM_SUBTLE_TRAIN = '/path/to/PRISM/Subtle/imgs'
PRISM_FULL_TRAIN   = '/path/to/PRISM/Full/imgs'

---

Training

Two-phase training: phase 1 trains only the classification head on PRISM_Full with an energy-based loss; phase 2 fine-tunes the full network on PRISM_Subtle adding a triplet contrastive objective.

python train.py \
  --save_dir ./Results \
  --save_name my_run \
  --device cuda:0

Key arguments (defaults match the best model):

Argument	Default	Description
--resolution	384	Input resolution
--lr1	1e-3	Learning rate, phase 1
--lr2	1e-6	Learning rate, phase 2
--lambda_w	0.3	Triplet loss weight (0 = BCE only)
--margin	0.1	Triplet loss margin
--training_patience	7	Early stopping patience, phase 1
--fine_tuning_patience	7	Early stopping patience, phase 2
--comp_prob	0.2	JPEG augmentation probability
--min_quality	50	Minimum JPEG quality for augmentation
--device	cuda:1	Device to train on

---

Evaluation

python evaluate.py \
  --model_dir /path/to/model_folder \
  --real_dir  /path/to/real/images \
  --generated_dir /path/to/generated/images \
  --device cuda:0

---

Pretrained detector

You can download the pretrained checkpoints from the link below:

• Checkpoints

---

Citation

@article{BARTOLUCCI2026104826,
  title   = {The PRISM benchmark: PhotoRealistic Image Synthesis and Manipulation to detect generated images},
  journal = {Computer Vision and Image Understanding},
  pages   = {104826},
  year    = {2026},
  doi     = {10.1016/j.cviu.2026.104826},
  url     = {https://www.sciencedirect.com/science/article/pii/S1077314226001931},
  author  = {Filippo Bartolucci and Samuele Salti and Giuseppe Lisanti}
}