Identifying Prompted Artist Names from Generated Images

Grace Su, Sheng-Yu Wang, Aaron Hertzmann, Eli Shechtman, Jun-Yan Zhu, Richard Zhang
arXiv, 2025

Website | Paper | Dataset

Abstract

A common and controversial use of text-to-image models is to generate pictures by explicitly naming artists, such as “in the style of Greg Rutkowski”. We introduce a benchmark for prompted-artist recognition: predicting which artist names were invoked in the prompt from the image alone. The dataset contains 1.95M images covering 110 artists and spans four generalization settings: held-out artists, increasing prompt complexity, multiple-artist prompts, and different text-to-image models. We evaluate feature similarity baselines, contrastive style descriptors, data attribution methods, supervised classifiers, and few-shot prototypical networks. Generalization patterns vary: supervised and few-shot models excel on seen artists and complex prompts, whereas style descriptors transfer better when the artist’s style is pronounced; multi-artist prompts remain the most challenging. Our benchmark reveals substantial headroom and provides a public testbed to advance the responsible moderation of text-to-image models. We release the dataset and benchmark to foster further research.

To Do:

• Dataset generation code

Getting Started

git clone https://github.com/graceduansu/IdentifyingPromptedArtists.git
cd IdentifyingPromptedArtists
conda env create --file environment.yaml
conda activate ipa

Download the Dataset

(See the HuggingFace dataset page for explanations of different data splits and data fields.)
First, login to HuggingFace using your token to download from this gated dataset:

huggingface-cli login

If your machine has high network bandwith, you can enable hf_transfer to speed up the download:

pip install huggingface_hub[hf_transfer]
export HF_HUB_ENABLE_HF_TRANSFER=1

Sample Dataset

The entire benchmark dataset consists of 1.95 million images and is ~2.9TB in size. To facilitate quick testing, we provide a sample dataset of 17 GB that is structured similarly to the full dataset. The sample dataset contains 10,200 SDXL images, with half of the images generated with complex prompts and half with simple prompts.

To download the sample dataset to the current directory and extract all images, run:

python scripts/download_dataset.py --download_config sample

Full Dataset

To download the full dataset to the current directory and extract all images, run:

python scripts/download_dataset.py --download_config all

--download_config can be sample, single_artist, multi_artist, all, or laion. See scripts/download_dataset.py parse_args() for more details and download options.

Dataset Location and Structure

All dataset files will be downloaded to the current directory.
After downloading, the repository will be structured as follows:

IdentifyingPromptedArtists
├── dataset_all_sample_sdxl
│   ├── train_imgs.csv
│   ├── test_artist_imgs.csv
│   ├── ...
├── dataset_all_single_artist
│   ├── train_imgs.csv
│   ├── test_artist_imgs.csv
│   ├── ...
├── sdxl_images_easy_prompts-artist
│   ├── abraham_pether_seed0
│   │   ├── xxx.png
│   │   ├── ...
│   ├── abraham_pether_seed1
│   │   ├── xxx.png
│   │   ├── ...
│   ├── ...
├── sdxl_images_mj_prompts-artist
│   ├── abraham_pether_seed0
│   │   ├── xxx.png
│   │   ├── ...
│   ├── abraham_pether_seed1
│   │   ├── xxx.png
│   │   ├── ...
│   ├── ...

Directories with the dataset_ prefix contain different subsets of the dataset. Each directory contains a CSV file for each of the subset's splits (train, test_artist, test_all_unseen_query, test_all_unseen_support). Each CSV file contains the image paths, labels, and train/test splits for different subsets.

Directories matching *_images_* contain the generated images for different text-to-image models and prompt types.

Download Trained Models (optional)

Our models

We provide the model checkpoints for the vanilla classifier and prototypical networks trained on our dataset.
To download the model checkpoints to the current directory, run:

python scripts/download_models.py --download_config [DOWNLOAD_CONFIG]

The model checkpoints will be saved to the trained_models/ sub-directory.
If prototypical networks are downloaded, the seen and held-out artist prototype features will also be downloaded and saved to the dataset_laion and dataset_laion_for_journeydb sub-directories (Each prototype is saved as a .npy file inside the dataset directory used to compute it).

The download_config options are:

• all: Download all models
• single_artist_prototypical: Download the prototypical network trained on all single-artist prompted images
• multi_artist_prototypical: Download the prototypical network trained with a multi-label objective on all SDXL single-artist and multi-artist prompted images
• single_artist_vanilla: Download the vanilla classifier trained on single-artist prompted images

Baselines evaluated in the paper

Please see https://github.com/learn2phoenix/CSD?tab=readme-ov-file for instructions on downloading the CSD (Contrastive Style Descriptors) model.

Please see https://github.com/peterwang512/GenDataAttribution for instructions on downloading the AbC (Attribution by Customization) models.

Model Evaluation on Benchmark

Please see README_evaluation.md for instructions on how to evaluate different models on our benchmark.

Model Training

Please see README_training.md for instructions on how to train the prototypical network and vanilla classifier on our dataset.

Acknowledgements

Our code and dataset is built upon the following works. We are grateful to the authors for sharing their code/dataset/model.

• Contrastive Style Descriptors (https://github.com/learn2phoenix/CSD)
• JourneyDB (https://journeydb.github.io/)
• PixArt-Sigma (https://pixart-alpha.github.io/PixArt-sigma-project/)

BibTeX Citation

@article{su2025identifying,
  title={Identifying Prompted Artist Names from Generated Images},
  author={Su, Grace and Wang, Sheng-Yu and Hertzmann, Aaron and Shechtman, Eli and Zhu, Jun-Yan and Zhang, Richard},
  journal={arXiv preprint arXiv:2507.18633},
  year={2025}
}