search

cure-lab / MMA-Diffusion

[CVPR2024] MMA-Diffusion: MultiModal Attack on Diffusion Models
Other
85 stars 5 forks source link

About the datasets #4

Closed xinqilyu closed 4 weeks ago

xinqilyu commented 4 weeks ago

Where are the text-image pairs from? In this paper, I see you write two datasets: LAION-COCO and UnsafeDiff. They both are text prompts. So i wonder where the clean original image are from, and their paired text prompt.

yangyijune commented 4 weeks ago

Hi! The clean images were collected from the Midjourney website's showcase gallery (https://www.midjourney.com/showcase) and are available in our Hugging Face dataset (https://huggingface.co/datasets/YijunYang280/MMA_Diffusion_adv_images_benchmark). However, we did not collect the original prompts associated with these images.

xinqilyu commented 4 weeks ago

so in your paper, how you match the text prompt "A completely naked Trump stands on the grass" with the image of Trump?

yangyijune commented 4 weeks ago

The images in Figures 2 and 4 are real photos collected from the internet for illustrative purposes. "A completely naked Trump stands on the grass" was written by me.