Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Search 5.8B images used to train popular AI art models

haveibeentrained.com

Post reply on HN
Print view View original

71–80 of 187 posts

Re: Search 5.8B images used to train popular AI art models

#71

Post by hammyhavoc »

dQw4w9WgXcQ wrote: post #70
hammyhavoc wrote: post #68

Earlier quoted context omitted.

I haven't learned from any photographers. https://people.com/music/pharell-williams-robin-thicke-blurr...

I'm impressed, never looked at another photo, just emerged from a blank room taking pictures having never seen another photo before. Just knew photography from the womb and came out with camera in hand knowing exactly how to work it. You are the one man who never had to learn a thing or rely on someone else or their prior works for training and/or inspiration.

I've been playing with DSLRs for twenty years. I just fuck around until I have some rudimentary understanding of it. Just like how I approach anything else, including playing instruments.

Re: Search 5.8B images used to train popular AI art models

#72

Post by egypturnash »

XorNot wrote: post #39
hammyhavoc wrote: post #33

Wow, so I've been trained. Many times over. Photos stolen from my personal portfolio.

You mean the portfolio which, by definition to be in that dataset, was marked for indexing by robots.txt, and made available publicly to unauthenticated GET requests on the internet?

Amazingly enough nobody made choices about how public they wanted their work to be that anticipated "someone scraping the entire internet and putting it into a giant neural network". This whole AI training thing may technically fall under fair use but it is right on the edge and begging for this very hazy edge to be made a lot sharper.

Re: Search 5.8B images used to train popular AI art models

#73

Post by dQw4w9WgXcQ »

hammyhavoc wrote: post #71
dQw4w9WgXcQ wrote: post #70

Earlier quoted context omitted.

I'm impressed, never looked at another photo, just emerged from a blank room taking pictures having never seen another photo before. Just knew photography from the womb and came out with camera in hand knowing exactly how to work it. You are the one man who never had to learn a thing or rely on someone else or their prior works for training and/or inspiration.

I've been playing with DSLRs for twenty years. I just fuck around until I have some rudimentary understanding of it. Just like how I approach anything else, including playing instruments.

Yea, the only "mental gymnastics" here is the BS you're trying to pass us that your personal neural network has absorbed exactly zero input or observation from other creators/creations that influences (consciously or otherwise) it's ability to create and profit from works.

Re: Search 5.8B images used to train popular AI art models

#74

Post by XorNot »

0xf00ff00f wrote: post #66
XorNot wrote: post #63

Earlier quoted context omitted.

So you're trying to assert copyright on something which isn't a copy of something? Again, where in the neural weights does your original work live? EDIT: Let's consider a simpler scenario - I take an MD5 sum of one of your photos, and then hash collide a an image from my phone camera till I find a match. Which part of this process is "stealing"?

If I compress an image with JPEG, I won't get the exact same pixels values in the original image after decompressing because the compression is lossy, but I'm sure you'll agree that the original image is "there". How are NN weights different from the DCT coefficients in a JPEG file?

The NN weights can't recreate anything without input values - specifically a 512x512 grid of random noise, and a transformed textual prompt.

So there's almost a kilobyte of missing data, without which nothing is produced.

Would a JPEG file with random noise in it be potentially any given original image? Of course not - even though the decompressor is perfectly capable of recreating one given suitable input data.

The "Grokking Stable Diffusion" colab posted here a week or two back makes this particularly explicit - any given 512x512 image can be reversed into a latent-space encoding of that image, and reconstructed from it. The NN weights are necessary but not sufficient to do so - ultimately you end up with a 512 byte number mapping to the expression which reconstructs an image. But that includes images which aren't part of the original training set.

[1] https://colab.research.google.com/drive/1dlgggNa5Mz8sEAGU0wF...

Re: Search 5.8B images used to train popular AI art models

#75

Post by Jach »

UncleEntity wrote: post #49
quantified wrote: post #41

Earlier quoted context omitted.

I'll get downvoted massively, but porn is by definition art (meant to be viewed and evoke reactions, just as advertising art is), and has been some of the most important creative outputs for as almost as long as humans have been creating art.

> meant to be viewed and evoke reactions, just as advertising art is And when they mix the two things get interesting. Having topless women selling (cow’s) milk was a very strange experience as a “prudish” American.

Sometimes I think a single change that would have an outsized improvement on American culture would be to make the mandatory high school art class include a week of sketch-drawing naked models, as actual artists do, both male and female.

Re: Search 5.8B images used to train popular AI art models

#76

Post by Jach »

simandl wrote: post #31
shantara wrote: post #27

Does this website search the same dataset as https://knn5.laion.ai ?

Yes, that's the same dataset. This website has some additional tools coming so artists can flag and opt out, or upload to opt in, and we'll get those to the laion team to add or remove from the 5B (and future) datasets.

What's your approach to establishing authorship? I'm sure some trolls will try to claim things they didn't actually make and get them removed (and vice versa).

Re: Search 5.8B images used to train popular AI art models

#77

Post by hammyhavoc »

dQw4w9WgXcQ wrote: post #73
hammyhavoc wrote: post #71

Earlier quoted context omitted.

I've been playing with DSLRs for twenty years. I just fuck around until I have some rudimentary understanding of it. Just like how I approach anything else, including playing instruments.

Yea, the only "mental gymnastics" here is the BS you're trying to pass us that your personal neural network has absorbed exactly zero input or observation from other creators/creations that influences (consciously or otherwise) it's ability to create and profit from works.

Strange ad hominem.

I capture what I see, more as a frame of reference for myself with location scouting and game dev. I take quite literally hundreds of thousands of photographs. It's 1 in every thousand that's worth sharing.

I have zero interest in other photographers, I can't even name one. Whatever other people do, neat. I don't even consider myself to be a "photographer", but I guess with hotels and airports buying prints, that somehow makes me a professional.

I capture what I like. There's literally zero other motivation behind it. Most of my income comes from elsewhere. I use a print-on-demand service.

If I take a photograph of a sculpture I've made and that is then incorporated into another work without permission, that's intellectual property theft. A lot of the images are of things I've made, not just street photography or landscapes that anybody could go and take.

If I write a riff in a microtonal scale and it turns up as a sample in a record, or is covered and released without permission, that's intellectual property theft too.

Re: Search 5.8B images used to train popular AI art models

#78

Post by MacsHeadroom »

hammyhavoc wrote: post #65
XorNot wrote: post #63

Earlier quoted context omitted.

So you're trying to assert copyright on something which isn't a copy of something? Again, where in the neural weights does your original work live? EDIT: Let's consider a simpler scenario - I take an MD5 sum of one of your photos, and then hash collide a an image from my phone camera till I find a match. Which part of this process is "stealing"?

The part where you used my work commercially. Seeing as you like analogies: if you play music in a public building, you need a license even if you are selling coffee and not the music.

Assuming you're referring to Stable Diffusion, the training set photos are not used commercially. Using them in training is non-commercial. If they were to share the training set, that would be commercial. However, the final product (the neural weights) does not include any of the training data and so there is no commercial use of your photos occuring.

It's the same as if I were to look at your photos and then take my own similar photos for the background of a web store. I would have "used" your photos in some sense. But I would not have used them commercially in any way whatsoever.

You have just as much grounds for complaint in Stability's case as you would in the case of me looking at your photos as reference for my own.

Re: Search 5.8B images used to train popular AI art models

#79

Post by simandl »

Jach wrote: post #76
simandl wrote: post #31

Earlier quoted context omitted.

Yes, that's the same dataset. This website has some additional tools coming so artists can flag and opt out, or upload to opt in, and we'll get those to the laion team to add or remove from the 5B (and future) datasets.

What's your approach to establishing authorship? I'm sure some trolls will try to claim things they didn't actually make and get them removed (and vice versa).

Our initial approach will be to validate the artists manually, and trust them to only flag or upload their own works. If we hit a scale where that starts to become an issue, we have plenty of ideas for mitigating this, but we're not optimizing for that problem yet.

Re: Search 5.8B images used to train popular AI art models

#80

Post by cercatrova »

hahajk wrote: post #52

This is great just from a prompt engineering perspective. Now I can see what labels are typically like for the types of images I'm looking for instead of guessing.

You can do that with lexica.art too, nothing special with this particular implementation that OP has.
Post reply on HN