[CVPR Findings 2026] PRADA: Probability-Ratio-Based Attribution and Detection of Autoregressive-Generated Images

Simon Damm, Jonas Ricker, Henning Petzka, and Asja Fischer (Ruhr University Bochum)
Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Findings, 2026

This repository contains the official implementation of PRADA and information on how to use the DARG dataset.

Setup

We recommend using uv. Once installed, you can recreate our environment by running

uv sync

from within the root directory. Depending on your environment, you might have to make changes to pyproject.toml to select the right PyTorch version.

The DARG Dataset (Dataset of AutoRegressive-Generated Images)

Generated Images

Our dataset can be downloaded from Zenodo. In total, It contains 155,000 images from 20 autoregressive image generators.

• Class-to-Image (10,000 images per model, 10 per ImageNet class)
  • HMAR-d16/-d20/-d24/-d30
  • LlamaGen-B/-L
  • RAR-B/-L/-XL/-XXL
  • VAR-d16/-d20/-d24/-d30/-d36
• Text-to-Image (1,000 images per model, prompts from Synthbuster, corresponding to real images from RAISE-1k)
  • Infinity-2B
  • Janus-1B/-7B
  • LlamaGen-XL
  • Switti-1024

Note: The evaluation in the paper uses a reduced subset.

Generating your own images: If you need more images or would like to reproduce our dataset, please refer to scripts/generate_images.py.

Real Images

• Download the validation set from the official ImageNet website and run uv run scripts/prepare_imagenet.py path/to/imagenet/val to recreate out subset of 10000 images.
• Download RAISE_1k.csv from the RAISE-1k website and run uv run scripts/prepare_raise1k.py path/to/RAISE_1k.csv to download and convert the images.

Reproducing Our Results

To reproduce our results for class-to-image models, run

uv run predict.py --real-image-dir data/darg/c2i/real --fake-image-dirs "['data/darg/c2i/hmar_d20','data/darg/c2i/hmar_d30','data/darg/c2i/llamagen_b256','data/darg/c2i/llamagen_l256','data/darg/c2i/rar_l','data/darg/c2i/rar_xxl','data/darg/c2i/var_d20','data/darg/c2i/var_d30']" --mode c2i --train-size 250 --output-root results/c2i

To reproduce our results for text-to-image models, run

uv run predict.py --real-image-dir data/darg/t2i/real --fake-image-dirs "['data/darg/t2i/infinity_2b','data/darg/t2i/janus_1b','data/darg/t2i/llamagen_xlstage2','data/darg/t2i/switti_1024']" --mode t2i --train-size 250 --output-root results/t2i

These commands will automatically use our pre-trained score functions.

External Dependencies

Code in src/external contains external repositories (with slight modifications for fixing imports) that are subject to their own licenses:

• HMAR
• LlamaGen
• VAR
• RAR
• Infinity
• Janus
• Switti

Citation

If you find this repository or the DARG dataset helpful, please cite our work as follows:

@InProceedings{Damm_2026_CVPR,
    author    = {Damm, Simon and Ricker, Jonas and Petzka, Henning and Fischer, Asja},
    title     = {PRADA: Probability-Ratio-Based Attribution and Detection of Autoregressive-Generated Images},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Findings},
    month     = {June},
    year      = {2026},
    pages     = {6506-6516}
}