README.md

---
dataset_info:
  features:
  - name: image
    dtype: image
  - name: prompt
    dtype: string
  - name: id
    dtype: string
  - name: promptid
    dtype: string
  - name: width
    dtype: uint16
  - name: height
    dtype: uint16
  - name: seed
    dtype: uint32
  - name: grid
    dtype: bool
  - name: model
    dtype: string
  - name: nsfw
    dtype: string
  - name: subject
    dtype: string
  - name: modifier10
    sequence: string
  - name: modifier10_vector
    sequence: float64
  splits:
  - name: train
    num_bytes: 7270597392.368
    num_examples: 49173
  - name: test
    num_bytes: 1765294302.142
    num_examples: 12294
  download_size: 5194348793
  dataset_size: 9035891694.51
license: cc-by-4.0
task_categories:
- text-to-image
- image-to-text
language:
- en
tags:
- prompts
- engineering
- research paper
pretty_name: LexicaDataset
size_categories:
- 10K<n<100K
---

## Dataset Description

- **Repository:** [Github repository](https://github.com/verazuo/prompt-stealing-attack)
- **Distribution:** [LexicaDataset on HuggingFace](https://huggingface.co/datasets/vera365/lexica_dataset)
- **Paper:** [Prompt Stealing Attacks Against Text-to-Image Generation Models](https://arxiv.org/abs/2302.09923)
- **Point of Contact:** [Xinyue Shen](xinyue.shen@cispa.de)

### Dataset Summary

LexicaDataset is a large-scale text-to-image prompt dataset shared in [[USENIX'24] Prompt Stealing Attacks Against Text-to-Image Generation Models](https://arxiv.org/abs/2302.09923).
It contains **61,467 prompt-image pairs** collected from [Lexica](https://lexica.art/).
All prompts are curated by real users and images are generated by Stable Diffusion.
Data collection details can be found in the paper.

### Data Splits

We randomly sample 80% of a dataset as the training dataset and the rest 20% as the testing dataset.


### Loading LexicaDataset


You can use the Hugging Face [`Datasets`](https://huggingface.co/docs/datasets/quickstart) library to easily load prompts and images from LexicaDataset.

```python
import numpy as np
from datasets import load_dataset

trainset = load_dataset('vera365/lexica_dataset', split='train')
testset  = load_dataset('vera365/lexica_dataset', split='test')
```

### Metadata Schema

`trainset` and `testset` share the same schema.

| Column              | Type       | Description                                                  |
| :------------------ | :--------- | :----------------------------------------------------------- |
| `image`             | `image`    | The generated image                                          |
| `prompt`            | `string`   | The text prompt used to generate this image                  |
| `id`                | `string`   | Image UUID                                                   |
| `promptid`          | `string`   | Prompt UUID                                                  |
| `width`             | `uint16`   | Image width                                                  |
| `height`            | `uint16`   | Image height                                                 |
| `seed`              | `uint32`   | Random seed used to generate this image.                     |
| `grid`              | `bool`     | Whether the image is composed of multiple smaller images arranged in a grid |
| `model`             | `string`   | Model used to generate the image                             |
| `nsfw`              | `string`   | Whether the image is NSFW                                    |
| `subject`           | `string`   | the subject/object depicted in the image, extracted from the prompt |
| `modifier10`        | `sequence` | Modifiers in the prompt that appear more than 10 times in the whole dataset. We regard them as labels to train the modifier detector |
| `modifier10_vector` | `sequence` | One-hot vector of `modifier10`                               |


## Ethics & Disclosure

According to the [terms and conditions of Lexica](https://lexica.art/terms), images on the website are available under the Creative Commons Noncommercial 4.0 Attribution International License. We strictly followed Lexica’s Terms and Conditions, utilized only the official Lexica API for data retrieval, and disclosed our research to Lexica. We also responsibly disclosed our findings to related prompt marketplaces.

## Licensing

The LexicaDataset dataset is available under the [CC-BY 4.0 License](https://creativecommons.org/licenses/by/4.0/).

## Citation

If you find this useful in your research, please consider citing:

```bibtex
@inproceedings{SQBZ24,
  author = {Xinyue Shen and Yiting Qu and Michael Backes and Yang Zhang},
  title = {{Prompt Stealing Attacks Against Text-to-Image Generation Models}},
  booktitle = {{USENIX Security Symposium (USENIX Security)}},
  publisher = {USENIX},
  year = {2024}
}
```