Published August 17, 2023 | Version v1
Dataset Open

Unsafe Diffusion: On the Generation of Unsafe Images and Hateful Memes From Text-To-Image Models

  • 1. CISPA Helmholtz Center for Information Security
  • 2. Delft University of Technology, Netherlands

Description

[Update] Looking for a larger unsafe image dataset? We publish a new dataset named UnsafeBench on Hugging Face. Take a look at here!

This dataset used in the paper  https://arxiv.org/pdf/2305.13873.pdf contains four prompt sets and one image set.

The four prompt sets were used to query Text-to-Image models and generate images for safety assessment. These sets include three harmful prompt sets and one harmless prompt set. The harmful prompts originate from different sources and contain various unsafe concepts, such as sexually explicit, violent, disturbing, hateful, and political content.

Prompt Sets:

  • 4chan Prompts: Harmful
  • Lexica Prompts: Harmful
  • Template Prompts: Harmful
  • COCO Prompts: Harmless

Image Dataset:

This dataset consists of 800 images, which were randomly selected from all the generated images from Text-to-Image models.

  • Safe: 580 images
  • Sexually Explicit: 48 images
  • Violent: 45 images
  • Disturbing: 68 images
  • Hateful: 35 images
  • Political: 50 images

Files

Unsafe Prompts&Images Dataset.zip

Files (210.2 MB)

Name Size Download all
md5:cbe1b381fb9d749a8c36c095eb99ea75
210.2 MB Preview Download