PRISM-36K: A Benchmark Dataset for AI-Generated Image Attribution

PRISM-36K is a benchmark dataset of 36,000 AI-generated images for model-attribution research — the task of identifying which generative model
produced a given image.
It accompanies the paper "PRISM: Phase-enhanced Radial-based Image Signature Mapping for AI-Generated Image Attribution" (Ricco, Onofri, Cima, Cresci, Di Pietro; arXiv:2509.15270).

What is in the dataset

The dataset contains 36,000 PNG images at 512 × 512 pixels, balanced across six text-to-image generators with 6,000 images per model:

• DALL-E 2 (Ramesh et al., 2022) — closed, accessed via OpenAI API
• FuseDream (Liu et al., 2021) — GAN + CLIP guidance
• PixArt-α (Chen et al., 2024) — diffusion transformer
• SANA (Xie et al., 2024) — diffusion transformer
• Stable Diffusion 1.4 (Rombach et al., 2022) — latent diffusion
• VQGAN-CLIP (Esser et al., 2021) — GAN + CLIP guidance

Each generator produces 150 images per prompt over a fixed set of 40 author-written English prompts (20 short + 20 long, paired by topic).
All images are stored in lossless PNG format to preserve frequency-domain artefacts that are critical to spectral attribution methods.

What makes this dataset useful

• Prompt-matched generations. The same 40 prompts are issued to every generator, so cross-model differences reflect generator-specific signatures rather than prompt drift.
• Architectural diversity. The six generators span GAN-based, CLIP-guided, and transformer-based diffusion families, with both open-weight and closed-API systems represented.
• Reproducible splits. 100 random prompt-level train/test splits used in the paper are shipped as splits/splits_100.csv; one canonical "average split" (splits/average_split.json) is provided for direct reproduction of all figures and tables.
• Lossless integrity. Every image ships with a SHA-256 hash in checksums/SHA256SUMS (BSD-style, compatible with sha256sum -c) so users can verify their downloads.
• Rich metadata. Per-image manifest (metadata/images.csv) and prompt manifest (metadata/prompts.csv) support filtering by model, prompt length, prompt pair, or specific generation iteration.

Repository layout

PRISM-36K/
├── README.md
├── LICENSE.txt
├── CITATION.cff
├── CHANGELOG.md
├── metadata/
│   ├── prompts.csv
│   └── images.csv
├── splits/
│   ├── average_split.json
│   └── splits_100.csv
├── images/
│   ├── DALLE-2/
│   ├── FuseDream/
│   ├── PixArt-alpha/
│   ├── SANA/
│   ├── StableDiffusion-1.4/
│   └── VQGAN-CLIP/
└── checksums/
    └── SHA256SUMS

Image filename convention: <ModelName>_<promptid>_<iter>.png, with promptid ∈ 1..40 and iter ∈ 1..150.

Intended uses

• Training and evaluating model-attribution classifiers for AI-generated images.
• Benchmarking real vs. fake detectors in a controlled multi-source setting.
• Studying frequency-domain and spectral fingerprints of generative models.
• Research on content provenance, generative-AI accountability, and related forensic problems.

Companion resources

• Paper: arXiv:2509.15270
• Image-generation scripts (the code used to produce these images): github.com/emarich/PRISM-36K
• PRISM classifier and evaluation code: released upon full paper acceptance.

Licensing

Dataset (images and metadata): Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0).

Note on DALL-E 2 images. The 6,000 images in images/dalle2/ were generated via OpenAI's paid API and are subject to OpenAI's usage policies in addition to CC BY 4.0: users intending to use these images beyond academic research should consult OpenAI's current terms of service.

Note on NVIDIA-SANA images. The 6,000 images in images/sana/ are licensed under the Apache License 2.0 usage policies in addition to CC BY 4.0.

Citing PRISM-36K

If you use this dataset, please cite both the paper and this Zenodo record. BibTeX entries and a CFF citation file are provided in the repository (README.md, CITATION.cff).

Limitations

• Closed-set scope. The dataset covers six specific generators; it is not designed to support open-set attribution to unseen models.
• English-only prompts authored by the dataset creators; no multilingual or in-the-wild prompts are included.
• Synthetic only. No real photographs are included; for real vs. fake benchmarks, real images must be sourced from a complementary dataset.
• No identifiable individuals. Prompts were authored to elicit generic scenes (objects, animals, landscapes); the dataset contains no images of identifiable real persons by design.