Dataset · YijunYang280
Image generation filter bypass prompt benchmark
YijunYang280/MMA-Diffusion-NSFW-adv-prompts-benchmark
1,000 English text prompts for testing text-to-image model defense against NSFW attacks.
What's inside
Contains 1,000 English text prompts created using adversarial attack methods from the CVPR 2024 MMA-Diffusion paper. These queries can bypass built-in safety filters during image generation. The dataset is designed for developing and evaluating defense mechanisms in text-to-image models.
What you can do with it
- Evaluating safety filter robustness in text-to-image generators
- Training classifiers to detect hidden malicious prompts
- Tasks
- Text → image
- Languages
- English
- Size
- 1K–10K rows
- Downloads this month
- 228
- Likes
- 32
- Access
- on request on Hugging Face
How to load it
from datasets import load_dataset
ds = load_dataset("YijunYang280/MMA-Diffusion-NSFW-adv-prompts-benchmark")