Stable Diffusion is a big deal • Hacker News Board

Re: Stable Diffusion is a big deal

#341

Post by peoplefromibiza » Wed, Aug 31, 2022, 2:45 PM UTC

hbn wrote: post #331

peoplefromibiza wrote: post #326

Earlier quoted context omitted.

> not any more-so than I stole from him when I looked at that image I spoke with the author, the author disagrees with you. The reason he consider that stealing is because they took that image from the online shop where the disclaimer specifically states "2022 © Magnetic Press LLC by Arenathemes. All Rights Reserved" without even trying to contact the copyright holder. https://store.magnetic-press.com/products/viewpo…

It seems like you didn't understand my point > took that image from the online shop They didn't take the image. They looked at it and learned a few datapoints from it. It's not like they compressed the image and added it into the model, able to be reversed by anyone who downloads it. In fact it's likely there's other completely distinct images that could have contributed the exact same tweaks to the model, like a has…

> They didn't take the image. They looked at it and learned a few datapoints from it

Nobody authorized them to do what they did.

> No, but the analogy would be if you learned something from the blurb you read - like a word, or a factoid, or an expression - and it influenced some writing you did later. You didn't steal from that book's author

that's what clean room implementation is for

https://en.wikipedia.org/wiki/Clean_room_design

Wine contributors were not allowed to see Windows leaked source code to avoid copyright infringement claims.

both WINE and ReactOS have refused to use the leaks; ReactOS doesn't even allow people who have worked legitimately at MS in the past to be developers, simply because even the smell of contamination would expose the projects to enormous legal risks

So, to be fair, Stable Diffusion authors should take inspiration from all the source material, replicate it the best way they can using their own abilities and then train the model on what they produced.

They didn't do it for two reasons:

- it would have required centuries

- the result would have been much less compelling

stable diffusion is so interesting due to the fact that the source material was of high quality.

So, we can conclude that it's the merit of the authors of the source material, not simply the model itself.

Shouldn't they be rewarded or at least consulted before using their works?

What if some of them is training their own model on works they have the right to use and now Stable Diffusion took all their hard work away?

If stable diffusion was trained on my drawings, it would produce a steaming pile of shit.

And nobody would be talking about it right now...

Re: Stable Diffusion is a big deal

#342

Post by lelag » Wed, Aug 31, 2022, 2:46 PM UTC

pclmulqdq wrote: post #330

_gabe_ wrote: post #307

Is it just me, or do the comments in this thread seem to be the exact opposite of the sentiment in the comments on similar Github Copilot threads? I just find it a bit ironic that programmers are irate about Github Copilot using their copyrighted material to train. However, if it's an ML model training off of copyrighted artists material, clearly its a transformative work. I just find the opposing sentiments for thes…

If you use my copyrighted code to train your brain and reproduce exact copies of functions I wrote, I have a problem with that. Same for art. Exact copying is the copyright problem. If you use my copyrighted (public) code to train your brain, and then produce new functions that do new things, there is no problem. Replace "brain" with "AI" and you get my position on this (and a lot of others' positions). It seems to b…

Indeed. A difference I see between the GPT-like language models and those image generation models is that it feels like the language models actually hold full copies of a lot of sample of the training dataset (hence their ability to recite existing content), whereas the image generation models clearly do not: Stable Diffusion is 4Gb something, yet it can draw anything.

That's the amazing part: the training dataset contains 5B images. Yet it distilled all of them in this mere 4Gb of data and can produce an infinite amount of content with those. It really feel like it learned how to draw, in the same sense a human learn and does not simply reproduce exact copy of what it already saw.

Re: Stable Diffusion is a big deal

#343

Post by criddell » Wed, Aug 31, 2022, 2:49 PM UTC

LewisVerstappen wrote: post #5

imgabe wrote: post #4

Earlier quoted context omitted.

Fair use is a defense. You can still be sued for fair use, you have to go to court and prove to a judge or jury that your work constitutes fair use.

Yes but we have a common law system and there's already tons of precedent that training AI systems is transformative. It's also quite obvious just by looking at the generated images that it's clearly transformative. The images generated are unique and you can't trace the original copyrighted image from what's generated. You really don't need a judge to see that Fair Use covers Stable Diffusion.

> The images generated are unique and you can't trace the original copyrighted image from what's generated.

That might not always be true. I've gotten some results back that had the Getty watermark on them and others with the artist's signature. Unless the AI is adding that to images that never had one before (which might be a trademark issue), then you might be able to determine the provenance of the image components.

Is there something equivalent to the yellow dots printers add to their output that would survive the AI transformation?

Re: Stable Diffusion is a big deal

#344

Post by dagmx » Wed, Aug 31, 2022, 2:50 PM UTC

_gabe_ wrote: post #307

Is it just me, or do the comments in this thread seem to be the exact opposite of the sentiment in the comments on similar Github Copilot threads? I just find it a bit ironic that programmers are irate about Github Copilot using their copyrighted material to train. However, if it's an ML model training off of copyrighted artists material, clearly its a transformative work. I just find the opposing sentiments for thes…

That’s entirely how I feel too.

I think it’s either

A) almost like a dunning Kruger effect where people are less familiar with a domain and therefore think this has superseded it and cheer it on, whereas they’re more familiar with their own craft and can see where it falls short.

B) They see it as technology conquering a new area , but get nervous when it starts infringing on their own.

I think it’s a bit of both in all likelihood

Re: Stable Diffusion is a big deal

#345

Post by majou » Wed, Aug 31, 2022, 2:50 PM UTC

lelag wrote: post #342

pclmulqdq wrote: post #330

Earlier quoted context omitted.

If you use my copyrighted code to train your brain and reproduce exact copies of functions I wrote, I have a problem with that. Same for art. Exact copying is the copyright problem. If you use my copyrighted (public) code to train your brain, and then produce new functions that do new things, there is no problem. Replace "brain" with "AI" and you get my position on this (and a lot of others' positions). It seems to b…

Indeed. A difference I see between the GPT-like language models and those image generation models is that it feels like the language models actually hold full copies of a lot of sample of the training dataset (hence their ability to recite existing content), whereas the image generation models clearly do not: Stable Diffusion is 4Gb something, yet it can draw anything. That's the amazing part: the training dataset co…

32Gb (4GB)

Re: Stable Diffusion is a big deal

#346

Post by thunderbird120 » Wed, Aug 31, 2022, 2:50 PM UTC

pclmulqdq wrote: post #330

_gabe_ wrote: post #307

Is it just me, or do the comments in this thread seem to be the exact opposite of the sentiment in the comments on similar Github Copilot threads? I just find it a bit ironic that programmers are irate about Github Copilot using their copyrighted material to train. However, if it's an ML model training off of copyrighted artists material, clearly its a transformative work. I just find the opposing sentiments for thes…

If you use my copyrighted code to train your brain and reproduce exact copies of functions I wrote, I have a problem with that. Same for art. Exact copying is the copyright problem. If you use my copyrighted (public) code to train your brain, and then produce new functions that do new things, there is no problem. Replace "brain" with "AI" and you get my position on this (and a lot of others' positions). It seems to b…

Generative models attempt to model their training data. Essentially, they try to be a model of the underlying data distribution from which all samples in the training data were drawn from. A model of that distribution which cannot reproduce all samples from the original training data given the right prompting/query/seed is an incomplete model by definition. If it can't reproduce all samples drawn from the original distribution then it clearly does not model the same distribution those samples were drawn from.

That said, this is very very different from just copying at a conceptual level. This is going to end up being a an interesting legal question going forward. I'm curious to see how it turns out.

Re: Stable Diffusion is a big deal

#347

Post by prolyxis » Wed, Aug 31, 2022, 2:51 PM UTC

belter wrote: post #293

dylan604 wrote: post #288

Earlier quoted context omitted.

>Code that's 98% correct is completely useless. Whoa! Really? Seems like modern software releases would be excited to achieve 98% with the incredible amount of bug fixes/patches released very quickly after the massive beta test known as release day.

You are looking at code like a production line. It's a semantic construction. Code 0.1% wrong, sends you to the Sun instead of the Moon, debits your account instead of crediting...

What's your metric for the "percentage the code is wrong"? Is it how many lines of code were wrong, or how many test cases the code fails?

Presumably if AI-generated code passes every test case, but would fail on edge cases that some human programmer(s) did not anticipate in their suite of tests, the humans potentially might have made similar coding mistakes as the AI if they had had to personally write the code.

Re: Stable Diffusion is a big deal

#348

Post by Karunamon » Wed, Aug 31, 2022, 2:55 PM UTC

_gabe_ wrote: post #307

Is it just me, or do the comments in this thread seem to be the exact opposite of the sentiment in the comments on similar Github Copilot threads? I just find it a bit ironic that programmers are irate about Github Copilot using their copyrighted material to train. However, if it's an ML model training off of copyrighted artists material, clearly its a transformative work. I just find the opposing sentiments for thes…

The difference appears to be that Github Copilot just reproduces, verbatim, snippets of code without actually doing anything transformative to it. You cannot describe a sort algorithm and have it spit out something other than a direct reproduction of someone else's implementation.

With these artwork models, they can emulate general styles, occasionally known characters or bits of text will show up in the output, but it is (as far as I've seen) never a 1 to 1 faithful reproduction of the training material with no changes made.

This is an important difference from a legal perspective, not just a moral one. Whether or not a use of copyrighted work is transformative is a big part of whether that use is fair or not.

Re: Stable Diffusion is a big deal

#349

Post by Tomis02 » Wed, Aug 31, 2022, 2:57 PM UTC

habosa wrote: post #121

These new models completely changed my mind about how much impact AI will have in my lifetime. They are the most impressive software achievements in decades and anyone who has a “meh” reaction will absolutely end up looking silly. I’m not optimistic that their impact will be positive though.

> They are the most impressive software achievements in decades and anyone who has a “meh” reaction will absolutely end up looking silly.

With the risk of looking silly, I declare "meh" once more, just as I "meh"-eh when GPT-3 came out.

The so-called "AI" is not fundamentally different from the AI of the 80s, it's just that now we have much better hardware. The main problem of the past AI winters still remains - the existing algorithms focus on statistical methods, which can be rather inexact. Imagine a nuclear plant controlled by an AI, or an airplane flown by a neural network. They completely lack reasoning capability and therefore you can't trust they will be able to adapt to unpredictable situations.

Re: Stable Diffusion is a big deal

#350

Post by quitit » Wed, Aug 31, 2022, 2:58 PM UTC

zulban wrote: post #280

quitit wrote: post #246

I've said this ad nauseam - but people who think this is going to kill an industry clearly have no idea of said industry. It's a fantastically great tool, and a very exciting space, but reducing the function of creatives to people who draw pretty pictures is staggeringly ignorant.

Live music didn't disappear but the industry took a huge permanent culling when any cheap venue could just play a track. The replacement doesn't have to be great to trash an industry. It just has to be good enough for most. This is.

It's not a good comparison.

A better comparison is like saying AI text generation bots will replace authors, or that AI drug testing will replace drug development - which clearly isn't the case. The core issue is that people with no idea what the creative fields offer are throwing their hat in with what they think is going to happen. People with experience are saying the opposite because they know better.

At best, this will remove those websites where you can pay $5 for "a designer" to make you "an image" - is this a loss though? Such things have never been a threat to creative fields.