search

jaehong31 / SAFREE

SAFREE: Training-Free and Adaptive Guard for Safe Text-to-Image and Video Generation (arXiv: 2410.12761)
https://safree-safe-t2i-t2v.github.io/
16 stars 1 forks source link

Question about the /datasets/nudity-ring-a-bell.csv #2

Open beotborry opened 3 days ago

beotborry commented 3 days ago

Hello!. I'm Juhyeon Park studying AI in SNU, South Korea.

First of all, thank you for sharing the wonderful work!

I have a question on the /dataset/nudity-ring-a-bell.csv you shared.

As the filename indicates, I thought that the csv file includes adversarial prompts to generate nudity images provided by Ring-A-Bell authors, but I have found that the prompt differ from the original one shared by the authors of Ring-A-Bell. (https://huggingface.co/datasets/Chia15/RingABell-Nudity)

Can you explain why there is a difference?

Thank you in advance.

jaehong31 commented 19 hours ago

Thanks for your interest in our work :)

I used the same Ring-a-bell prompt dataset from https://github.com/CharlesGong12/RECE (ECCV 24’).

RECE authors obtained this dataset with ChatGPT, following Appendix H of ring-a-bell paper. They instructed ChatGPT to generate prompt pairs with and without unsafe adjunct words such as "bare, exposed, topless, and undress".

They said that this is because the repo (https://github.com/chiayi-hsu/Ring-A-Bell) was open-sourced in February and its nudity prompt pairs were made available in March, but this work was done before March (ECCV).

In addition, we followed their dataset for Ring-a-bell but ran baselines ourselves, ensuring fair comparisons.

beotborry commented 12 hours ago

Got it, Thank you for your response!