I found this very similar to what you'd get with a google / bing / etc image search. Is that where this database comes from? I noticed there is a lot of "Shutterstock" watermarked stuff. And I also checked a few "adult" terms (large breasts etc) and found there is a lot of nude content. Only curious because I've seen lots of the generative models have some post-filtering for nudity, why don't they just clean it out of the training data if they're worried?
Search 5.8B images used to train popular AI art models
21–30 of 187 posts
Re: Search 5.8B images used to train popular AI art models
#22Re: Search 5.8B images used to train popular AI art models
#23It's surprising how poorly labeled these images are; who is curating this collection? Can't they crowd-source a proper labeling project - I wonder how much better things like Stable Diffusion would be if its training would include correct, complete labels for the images. I'm sure lots of folks would willingly spend a few minutes here and there to aid with the labeling if it means they get to enjoy the model for free.
part of our idea is that artists can help with labelling their own work
Re: Search 5.8B images used to train popular AI art models
#24One thing that sticks out to me is how so much of the the images in the collection have really terrible labels. I uncovered a large collection of pieces by an illustrator who was unsearchable by name, only via image upload. The reason they were unsearchable: the majority of this particular artist's images had labels that were all in the format of:
{username}'s profile image
Re: Search 5.8B images used to train popular AI art models
#25I'd love to know how this works. I entered my own name for the lols, and it returned mostly paintings of the Cape Winelands in South Africa where I grew up, which is pretty creepy.
that is crazy. Perhaps your family name is common in that region?
Re: Search 5.8B images used to train popular AI art models
#26I just uploaded a picture of my dog (Bichon Frise) and it showed me a bazillion nearly exact similar dogs. Why isn't this immediately being used as a missing persons (or pet) database service?
sometimes people are missing, sometimes they just have legitimate reasons to not want to be found.
Re: Search 5.8B images used to train popular AI art models
#27Does this website search the same dataset as https://knn5.laion.ai?
Re: Search 5.8B images used to train popular AI art models
#28interestingly, if you search for "127.0.0.1", it throws {"message":"Forbidden"} in the api response and the page says- Sorry, there was an error with your search. Please try a different request.
Same for "localhost" but "[::1]" works.
Re: Search 5.8B images used to train popular AI art models
#29I found this very similar to what you'd get with a google / bing / etc image search. Is that where this database comes from? I noticed there is a lot of "Shutterstock" watermarked stuff. And I also checked a few "adult" terms (large breasts etc) and found there is a lot of nude content. Only curious because I've seen lots of the generative models have some post-filtering for nudity, why don't they just clean it out o…
This is Laion-5B, https://laion.ai/blog/laion-5b/
It's built off of common crawl, so it probably does have a pretty representative sample from whatever the big image searches use.
Funny enough, the NSFW filter that laion built is turned on. Without it, it's... a lot. The NSFW stuff is done with a model, so you get a probability of NSFW out of it, and you can select a threshold.
If you set the threshold high, like 95% certainty that an image is nsfw to filter it, you get a bunch of false negatives, letting a ton of nsfw through. Set it too low, and you throw out stuff that isn't nsfw.
We (haveibeentrained) erred on the side of too high, so we wouldn't tell artists their work wasn't in there if it was. Tough trade-off there. Similar to using the dataset to train an AI model, where you might cut off useful images from training if you try and filter all the nsfw.
Re: Search 5.8B images used to train popular AI art models
#30One thing that sticks out to me is how so much of the the images in the collection have really terrible labels. I uncovered a large collection of pieces by an illustrator who was unsearchable by name, only via image upload. The reason they were unsearchable: the majority of this particular artist's images had labels that were all in the format of: {username}'s profile image
I'm not 100% on this, but I think a big portion of the captions for the images come from the alt-text, and are probably auto-generated by the sites they were scraped from.