Hacker News Reader

However, it looks like the model wasn't trained on NSFW keywords anyway, which means it is permanently impaired in that regard.

It was trained on LAION 5b, which is chock-full of NSFW material. Search for yourself: https://rom1504.github.io/clip-retrieval/

They filtered a lot of it out, it is a subset of LAION 5b according to undisclosed parameters: https://stability.ai/blog/stable-diffusion-announcement