3 ms·

My understanding is it was trained on this dataset: https://rom1504.github.io/clip-retrieval/?back=https%3A%2F%2Fknn5.laion.ai&index=laion5B&useMclip=false http

by lbotos 4y ago

My understanding is it was trained on this dataset: https://rom1504.github.io/clip-retrieval/?back=https%3A%2F%2Fknn5.laion.ai&index=laion5B&useMclip=false https://rom1504.github.io/clip-retrieval/?back=https%3A%2F%2... Which has a LOT of NSFW images in it. I suspect if you removed them from the training set it would go a long way to curb NSFW output but as you say people could easily train their own NSFW latent diffusion model.

kmeisthax 4y ago
My experience with Stable Diffusion is that it has a habit of tripping the NSFW filter - as in, it actually generated NSFW images - on prompts that were entirely innocuous. Would not be surprised if Stability has a huge "how do we get it to stop spitting out porn so easily" problem.