Most active commenters
  • bambax(3)

←back to thread

As an artist I am concerned about AI image generation

(twitter.com)
114 points valgaze | 15 comments | 14 Aug 22 17:33 UTC | HN request time: 0.806s | source | bottom
Show context
adamhi ◴[14 Aug 22 19:02 UTC] No.32461913[source]▶
I won't pretend that this isn't a troubling development for digital artists, maybe even existentially so. I hope not.

One thing that makes me a little hopeful is that every image I've generated with DALL-E 2, even the best ones, would require non-trivial work to make them "good".

There's always something wrong, and you can't tell the model "the hat should be tilted about 5 about degrees", or "the hands should not look like ghoulish pretzels, thanks".

There's also this fundamental limitation that the model can give you a thing that fits some criteria, but it has no concept of the relationships between elements in a composition, or why things are the way they are. It's never exactly right.

It's like the model gets you the first 90%, and then you need a trained painter to get the second 90%.

But yeah, it will certainly devalue the craft, don't get me wrong. And anyone who is callously making comparisons to buggy whip manufacturers should consider how it would (excuse me, will) feel when AI code generators pivot to being more than a copilot, and suddenly the development team at your office is a lot smaller than it used to be, and maybe you aren't on it anymore.

If you spend a lifetime mastering some skill, and then it's just not valued anymore, it sucks, and you get pretty mad about it.

replies(6): >>32462125 #>>32462272 #>>32462281 #>>32462452 #>>32462520 #>>32463297 #
1. bambax ◴[14 Aug 22 20:13 UTC] No.32462520[source]▶
> If you spend a lifetime mastering some skill, and then it's just not valued anymore, it sucks, and you get pretty mad about it.

That is absolutely not what the OP is complaining about. They're not saying that because AI is good, they won't find work. They are complaining that in training AI for art generation, builders took works from living artists, without consent from them, and that in so doing allowed generators to make new art in the style of said artists.

The example given is that Stable Diffusion even tries to reproduce logos/signatures of living artists.

If I produced a rubbish search engine that bore a malformed "gigggle" logo using Google colors, how long do you thing I would survive before being sued out of existence by an army of Google lawyers?

But that's exactly what many AI generators are doing here.

Edit: the first version of this comment confused Stable Diffusion with OpenAI, and stated that OpenAI was owned by Google. OpenAI has a strong partnership with Microsoft. Stable Diffusion is not OpenAI. Sorry for the errors.

replies(2): >>32462609 #>>32462806 #
2. Kiro ◴[14 Aug 22 20:24 UTC] No.32462609[source]▶
Not only is your rant about Google misplaced considering Dall-E is OpenAI, not Google, but the thread is also not complaining about Dall-E. It's about Stable Diffusion (https://stability.ai/blog/stable-diffusion-announcement) which is explicitly trained on working artists. That's why it tries to reproduce the logo.
replies(2): >>32462720 #>>32462803 #
3. v64 ◴[14 Aug 22 20:38 UTC] No.32462720[source]▶
Dall-E has also been trained on watermarked art. Here [1] [2] are some examples from images I've generated exhibiting that.

[1] https://ibb.co/Q86zDSw

[2] https://ibb.co/njvLMQ2

4. bambax ◴[14 Aug 22 20:47 UTC] No.32462803[source]▶
You're right of course! I don't know what I was thinking. Dall-e is MS. I will edit.
replies(2): >>32463120 #>>32463646 #
5. WASDx ◴[14 Aug 22 20:47 UTC] No.32462806[source]▶
This is frankly what humans have always done, learning and taking inspiration from other artists. Now we have made a machine that can do the same thing.

In the case of exact reproductions, we have copyright and IP laws.

replies(2): >>32462917 #>>32463500 #
6. ThisIsMyAltFace ◴[14 Aug 22 21:04 UTC] No.32462917[source]▶
No. This argument comes up over and over and over again and it is wrong.

These models are not learning or being inspired in the same sense humans are. The laws tgat apply to humans should not be applied to them.

replies(2): >>32464376 #>>32465309 #
7. alex_young ◴[14 Aug 22 21:30 UTC] No.32463120{3}[source]▶
Still wrong. DALL-E is Open AI
replies(1): >>32463173 #
8. bambax ◴[14 Aug 22 21:35 UTC] No.32463173{4}[source]▶
Yes, that's what the edited comment says. Dall-e is OpenAI; OpenAI has strong links with MS.
9. greysphere ◴[14 Aug 22 22:12 UTC] No.32463500[source]▶
It feels like a big stretch to consider an algorithm to be 'inspired'. Where are the bits that correspond to 'inspiration'? Seems like that would answer a lot of big questions in philosophy.
replies(1): >>32464429 #
10. Kiro ◴[14 Aug 22 22:27 UTC] No.32463646{3}[source]▶
Thanks! Good edit. For the record I agree with your comment but got distracted by the error.
11. cercatrova ◴[15 Aug 22 00:04 UTC] No.32464376{3}[source]▶
You're right, AI generated pictures should not be copyrighted, as is the case today. People should be free to mix and remix pictures via AI as much as they desire.
replies(1): >>32473444 #
12. krapp ◴[15 Aug 22 00:15 UTC] No.32464429{3}[source]▶
Where are the neurons that correspond to "inspiration?" It's algorithms all the way down.
replies(1): >>32480572 #
13. slowmovintarget ◴[15 Aug 22 02:43 UTC] No.32465309{3}[source]▶
Exactly. The relevant law is with regard to the use of artwork on the part of the people who feed the index. Artists should have a say on whether their work gets included in a training set, if their works are not public domain.
14. PaulsWallet ◴[15 Aug 22 18:28 UTC] No.32473444{4}[source]▶
This is where I imagine things are going to get into trouble because how are you going to determine what is AI and what isn't? Especially when Stable Diffusion is directly classifying artist and cloning their signatures and watermarks. What about things that are started with AI and refined by human?
15. greysphere ◴[16 Aug 22 08:41 UTC] No.32480572{4}[source]▶
I claim that claiming computer algorithms are inspired is a big stretch.

I claim humans can be inspired.

I don't claim to know how human inspiration happens, or if neurons have anything to do with it. (They may, but I make no claim). Not being able to describe the process by which human inspiration happens doesn't invalidate either of my claims.

If there is a satisfactory non-bit based explanation to how computer algorithms achieve inspiration, I would accept that to. We have the advantage with computers, that their activity is conveniently summarized by their programs which are represented in bits, so expecting an explanation in that form I think is reasonable.

The defense if the claim of human inspiration is (1) we have that word for the concept (2) we have thousands of years of thought, philosophy and literature giving support and definition to the concept.