AI Art Generators Can Be Fooled into Making NSFW Images
spectrum.ieee.org
AI Art Generators Can Be Fooled into Making NSFW Images
1–10 of 193 posts
Re: AI Art Generators Can Be Fooled into Making NSFW Images
#2With or without humans providing AI rules/training set for what constitutes NSFW images?
Note: training set(s) would be extremely social/culture specific for given point in time/history.
Re: AI Art Generators Can Be Fooled into Making NSFW Images
#3Their GitHub link is 404 so here it is https://github.com/Yuchen413/text2image_safety
Re: AI Art Generators Can Be Fooled into Making NSFW Images
#4Is it still art, if it cant offend anyone anymore?
Re: AI Art Generators Can Be Fooled into Making NSFW Images
#5Their GitHub link is 404 so here it is https://github.com/Yuchen413/text2image_safety
Odd the repository has "safety" in the name. Are NSFW images considered "unsafe" in some weird way?
Re: AI Art Generators Can Be Fooled into Making NSFW Images
#6Is it still art, if it cant offend anyone anymore?
If it afflicts the comfortable or comforts the afflicted, it's art. If it does both, it's Fine Art.
Re: AI Art Generators Can Be Fooled into Making NSFW Images
#7I always assumed that image generation services had prompt and output filtering. The reason I believe this is that I've crafted prompts that seemed to pass the prompt filter, generated, and then were unable to be displayed[1]. Presumably there is an output filter that determines if the generation violates ToS against things like nudity. Though in my case, I wasn't able to generate images that too closely resembled Ansel Adams photographs. Shame on me, I guess.
1. This reads a little like being able to deduce what's going on by studying the timing, now that I reflect on it.
Re: AI Art Generators Can Be Fooled into Making NSFW Images
#8The root cause is that the text encoders (e.g. CLIPText) used to power image generation models are too good and can handle semantic synonyms, even oblique synonyms such as "Mickey Mouse" and "a famous anthropomorphic mouse mascot".
The repo mentioned focuses on NSFW content, but I've had an idea that can theoretically identify arbitrary cases that could be undesirable by image generation services: get a large dataset of CLIPText encoded texts, and train a LLM to take in said embeddings and output synonyms, or maybe something with contrastive loss like CLIP itself.