Lexica Dataset: 61,467 User-Curated Prompts and Stable Diffusion Images

61,467 prompt-image pairs collected from the Lexica platform for research on text-to-image generation models. All prompts were curated by real users and the corresponding images were generated by Stable Diffusion. The dataset was shared in a USENIX'24 paper and uploaded to Hugging Face by vera365 on May 16, 2024.

Use Cases

• Researching prompt stealing attacks on text-to-image models based on the dataset's origin in a USENIX'24 security paper.
• Analyzing the relationship between user-written prompts and AI-generated images based on the paired structure.
• Training or fine-tuning text-to-image models using a large-scale collection of real-world prompts.
• Studying user behavior and prompt styles from a curated online community source.

Strengths

• Contains 61,467 prompt-image pairs, providing a substantial corpus for analysis.
• Prompts are curated by real users, offering insights into real-world usage patterns.
• Images are generated by a specific, widely-used model (Stable Diffusion), ensuring consistency.
• Dataset is associated with peer-reviewed research (USENIX'24), indicating academic rigor.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is known, but specific data splits (80% train, 20% test) are described without downloadable files being confirmed.
• Data may reflect the biases inherent to the Lexica platform's user base and content.

Provenance

Source

Lexica platform.

Collection Method

Collected prompt-image pairs curated by users and generated by Stable Diffusion.

Time Range

null

Freshness

Last updated 2024-05-16 10:10:17.

Geography

null