cmu-gil/PromptedArtistIdentificationDataset

Prompted Artist Identification Dataset

Website | Paper | GitHub

Identifying Prompted Artist Names from Generated Images
Grace Su, Sheng-Yu Wang, Aaron Hertzmann, Eli Shechtman, Jun-Yan Zhu, Richard Zhang
arXiv, 2025

Prompted Artist Identification Benchmark. We introduce the first large-scale benchmark for identifying prompted artist names from generated images. The benchmark covers four axes of generalization that match realistic use cases: (1) Artists: we collect artists commonly used in prompts and simulate open-set artist classification by testing on artists not seen during training. (2) Prompt complexity: users describe images in many different ways, including short, simple prompts as well as more descriptive, complex prompts. (3) Text-to-image models: users can generate images using various text-to-image models, which have different training data and architectures that may affect the generated image’s overall style. (4) Number of artists: users may include multiple artists in the prompt to mix styles, creating images that are not easily attributable to a single artist.

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.

Browse and View Sample Dataset

The entire benchmark dataset consists of 1.95 million images and is ~2.9TB in size. Hence, the dataset viewer only contains text and does not display the images.

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 browse images in the sample dataset, visit the sample dataset's HuggingFace dataset viewer.

Download Instructions

The easiest way to download the dataset is to follow the instructions and use the script provided in the GitHub repository. Follow "Getting Started" and "Download the Dataset" in the README.md.

Dataset Structure

Subsets (Dataset Configs)

The dataset is divided into multiple subsets based on the text-to-image model used, prompt complexity, and the number of artists per prompt. All subsets can be viewed in the Dataset Viewer.

Data Splits

Each subset is further divided into the following data splits:

• train: Training images generated with seen artists.
• test_artist: Test images generated with seen artists and held-out prompts. Evaluates classification performance on seen artists.
• test_all_unseen_query: Test images generated with held-out artists and the remaining held-out prompts. Evaluates classification performance on held-out artists.
• test_all_unseen_support: Images to use as references during inference on test_all_unseen_query, generated with held-out artists and 5 held-out prompts.

Data Fields

• img_name: The image file path.
• prompt_label: The artist name used in the image's generation prompt.
• prompt_type: (Can be ignored) Whether the image invokes an artist name or is "style-only".
• source_label: The source of the image. Either "sdxl", "sd15", "pixart", or "midjourney".
• prompt_num: The ID of the prompt template used to generate the image. See metadata/map_complex_prompt_nums.txt and metadata/map_simple_prompt_nums.txt for the mapping of prompt IDs to text prompts.
• subject: If the image is generated with a simple prompt, the generation prompt is "A picture of [subject] in the style of [artist]".
• all_prompt_labels: A list of all artist names used in the prompt, separated by semicolons. This is useful for images generated with multiple artist names in the prompt.
• seed: The random seed used to generate the image.
• prompt_text: The full text prompt used to generate the image.

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}
}