Another AI vs artists thread

No worries, I hate it too!

If the training dataset had more precise geometric shapes, it is very likely it would be able to handle them a lot better. Not sure whether it would be good enough though. It is sort of like faces. Each tiny mistake is very obvious to us.

4 Likes

Here is a video about the legal situation in the US:

I really hope I’m wrong, but I’m afraid, taking into consideration what Midjourney is capable of now, that in the next few years 99% graphics jobs will be gone… concept artists, illustrators, mate painters, web designers, 3d artists… there will be only few of them, best of the best, who will stay in the industry, but others will be gone.
Who will hire an illustrator to make 30 illustrations for a book, if A.I. may do it in just few seconds, giving all these images in different variants and when any modification of it, is possible within another few seconds…
Many things will be done by people with the best language skills and artists will not be the only ones to lose their jobs…

This is what Midjourney created for me, based on just few words…


You clearly don’t work in web design/dev if you’re worried about AI taking over that field :sweat_smile: yeah, sure, making a good-looking website is easy and computers can do it no problem (already could long before Midjourney etc). What computers can’t do is make extremely specific revisions to keep up with an ever fickle client- or two clients, with one site, and conflicting interests, which is what happens on probably 80% of the sites I work on.

See, computers are logical, and clients aren’t. Most of the time, clients ask for revisions that look bad, or they need a custom PHP script to hook the contact form to some archaic database running on an original IBM tower they had the CEO’s nephew write in the 90s. AI can make things look good, humans can make things looks the way humans want- which isn’t good, and is insanely specific.

Show me an AI generator that can understand “I want classy and elegant, but also family friendly and whimsical, and it needs to match my anthropomorphic alligator logo” - not a hypothetical, I just finished that site :wink:

The reason I get paid every day is because I can take clients’ horrifying, disgusting, disturbing design/development requests and wrangle them into something professional. This isn’t a logical process, it requires an intense understanding of the client, and there’s no way to automate it. It’s radically different for every client and every project.

PS- my company has been using AI to write content for clients for the past five years, long before the current AI hype, so they’re not the “heads stuck in the sand” type. If what I’m describing could be automated, it would be already

1 Like

Well, as I already wrote, I hope you are right and I’m wrong…
Yes, I’m clearly a developer with over 20 years experience in Java, C#, Python, Go, Perl, Linux, Swift, JavaScript, Kotlin, iOS, Android, and so on, and so on.
And yes, I know how many fu…. up code exists in this world which no A.I. would understand.
Yes, I have worked also as web designer, that was long time, in the early 00’s, but I was tired to work with customers so I’ve focused purely on dev and I’m very happy not seeing any customers since then.
And yes, I could give many examples when dev is just a LEGO building process from existing components so I can imagine that A.I. can learn that if instructed correctly. Maybe not now but in few years definitely.
And I didn’t say that all of us will be jobless, I just think that with the assistance of A.I. one person will be able to do what now requires 10 persons.
So 9 out of 10 will be jobless not all :slight_smile:
And I know also how much effort is needed to make 3D similar to the images I’ve attached in the previous post. Skilled person would need 2-3 days to make one. Midjourney gave me 4 images in 60 seconds, and another 4 variants in next 60 seconds and I’m pretty sure that after few iterations the image would be perfect.
So yes. If machine is 3000x faster than human I’m afraid the same way I will be afraid when machines like terminator will walk the streets.

Interesting take on AI art:

Opinion on AI art legal issues from a relatively unbiased standpoint:

TLDW:

  1. Purely text-to-images almost certainly will never be able to get a copyright, though at what fraction of human intervention copyright is applicable remains to be seen.
  2. In the copyright infringement lawsuits there is a good chance the court will side with the AI companies. But this is an unprecedented situation, we don’t really know.
  3. He also talks about another interesting case where AI technology is used to commit intentional copyright infringement. Who is liable- the prompter, the original coder or nobody? Again, we don’t really know.

My personal opinion- even if the lawsuits are groundless, they are essential to set a precedent. Plus they are getting a lot of traction, informing the general public about the ethical issues at hand- letting them make an informed choice.

2 Likes

Where did you research them? Did you read books? Which ones?

There is a new study about memorization of diffusion models (here is the corresponding Twitter thread):

As far as I understand, there are obvious memorizations. Most of them because the images appeared many times in the dataset. Getting rid of them, should help according to the authors. However, even after that, some images were still memorized!
They are essentially exploring ways to better understand and reduce memorization.

Interesting study, but to me it shows the opposite of what they wanted.

  1. It takes a lot of effort and the extracted results are poor quality. In fact quote from the study itself

In contrast, we failed to identify any memo-
rization when applying the same methodology to Stable
Diffusion—even after attempting to extract the 10,000
most-outlier samples.

  1. So what? You can download these images from the internet, they are already public, so trying to extract them from an AI dataset is like circumnavigating the earth to get to your house. You could use a supercomputer to try and extract an image, or you could type it into google.

Their goal was to find out whether memorization takes place in those sorts of neural networks. And they could show that it happens, especially or mostly when data is redundant in the training dataset.
That’s important because memorization is in most cases not wanted, because the goal is to learn more general patterns which can be used in many situations. The more memorization there is, the less capacity there is for more general patterns to be learned.
Also, when you are able to extract copyrighted images from neural networks, there might be legal issues.

3 Likes

Yeah it is interesting that it happens. I guess maybe you could argue that they’re distributing copyrighted material by distributing the models that contain that material that can be extracted. But you can reverse engineer software too, usually you just agree not to do that when you download the software. It’ll be interesting to see if this has any impact.

But the point is when I distribute software, I need the rights to do that. It doesn’t matter whether users can reverse engineer the code.
On the other hand when I distribute neural networks that I trained with images I don’t have the copyright for, it shouldn’t be possible to extract the images I used for the training. After all, I don’t have the rights for those images.
The legal situation is not clear yet for those neural networks, but in my view, the situation is very different compared to reverse engineering.

2 Likes