Hacker News Reader

Comment by spaceman_2020 | Hacker News Reader

spaceman_2020

Old man rant incoming: Can someone tell me what's the need for AI art in the first place?

Like, why are we wasting resources to fix something that human beings actually enjoy doing themselves?

Shouldn't AI and robotics focus first on fixing issues that human beings don't like doing? Where's my robot that can chop onions for me?

Am I the only one who finds the current crop of AI startups a little misguided in their priorities? Instead of fixing unpleasant stuff in human lives, they're trying to takeover the fun and pleasant things we actually like to do - write, paint, code.

/endrant

View on HN

AI art is completely ripping off actual artists to build their learning datasets. They're datasets are full of copyrighted work but until now no one cares much about visual artists. Take a look at the statements for AI music generation and you'll see they're very careful to state they use non-copyrighted, royalty free source tracks. No such thing for artists

Beware the industry pushing AI art forwards, it's not at all an industry built for artists. You'll see "it's a tool for artists" or "it'll happen regardless" or "it won't get that good" or "it's useful for generating references/ideas," etc. Tell me what's the business model here?

They create non-profits to source the datasets to then pass to their for-profit parent company. I hope the class actions come quickly enough.

Tell me you're a neo-luddite without saying it... "ripping off"

The statement about non-copyrighted sources is because Google is majorly bad PR/conflict averse, so they're trying to sidestep all the "They're stealing my soul, they took my job, herp derp" rabble rousing.

I don't understand what your argument here is? Don't fight technology because it's 'the future?'

I enjoy making art, but I don't have the technical skill take the things I can visualize in my head commit them to a medium. AI art tools let me iterate using image transformation to get those things out.

I enjoy creating written stories, but I find the process of going from an outline/abbreviated rough draft to a polished story to be tedious. AI tools automating the process of finishing a written story would make the process of writing much more enjoyable.

I enjoy building software, but I don't enjoy writing a ton of boilerplate. AI writing the boilerplate for me would make the process of creating software much more enjoyable.

>I enjoy making art, but I don't have the technical skill take the things I can visualize in my head commit them to a medium. AI art tools let me iterate using image transformation to get those things out.

what do you mean by iterate here? do you mean getting an outline from a model to base your work off of, or just using multiple pass throughs of the same image to create a desired final product? as I would not call the latter the process of making art

My workflow involves photobashing to define the composition of the image, then a combination of img2img, inpainting and manual editing repeatedly until I achieve the desired results.

AI images with zero modifications have already had their copyright shot down, though that's effectively unenforceable unless someone freely admits the image was AI generated and not modified in any way.

I see this a lot, so I'll probably write a blog post on it sometime. But the short answer is that not everyone is good at art, but they know exactly what they want, and they want it as soon as possible.

There's no motive to harm artists, it's just that the tech is ready and there's a niche for it. Artists can actually benefit too, they can use this tech for inspiration and for a starting point, if nothing else.

Plus there are many use cases where it would be nice to have professionally made, unique art, but the costs of creating it would simply be too high.

For example, a group of friends sitting around playing D&D, would not likely commission custom artwork for each of their characters. With tools like stable diffusion, you could have a set of several images of each character, as well as for NPCs, landmarks, runes and so on, all in an afternoons work, with some knowledge of the tech.

Yes! AI-generated content has to work ok-ish to be useful.

I can imagine a game where depending on your level, mastery or expertise. You could create AI-generated gear system where depending on your stats you could get different styles.

> not everyone is good at art, but they know exactly what they want,

This is pretty much like saying "not everyone is good at programming, but they know exactly what they want". Then consider all the debate about agile processes and amount of ink spilled over how to get your customers to figure out what they want from your code...

Honestly, as much as I think AI art is deeply unfair to artists who unwillingly contributed its learning material and are getting replaced by "in the style of ..." prompts, I can see how just deferring a commercial customer to AI can be a blessing. Let the machine deal with "this looks fine and it fits the material I gave you, but I imagined it differently. Can you redo it by tomorrow?"

I worked in an agency. The way pretty much all creative process in any large company works is that the stakeholder (such as a brand manager) signs off on a broad vision for a creative campaign. The agency then uses its creative talent to create the actual collateral. Most of the times, the stakeholder will only be asked to look at (or want to look at) major campaign collateral (such as a superbowl ad).

All the stuff that AI is automating currently is too mundane and cheap to involve a decision maker. No brand manager is spending time signing off on 1 of 12 product images out of 2000 SKUs on their site.

My point is: the business use case for AI generated art is really overblown. Small businesses definitely interact with creative tools directly, and it will be of great help to them. But in any business of any significant size, the creative process is offloaded to the “creative types” and the only input decision makers will have is to tell them what something should look like. At no point will they actually sit down and try to use creative tools to create stuff.

General-purpose robotics is a much harder problem and costs more money.

Art turned out to be within reach of current ML techniques at a relatively reasonable price, so that’s where people are investing their resources right now.

It’s not that people consciously chose to automate art before automating manual labor. It just happened to work out that way.

The reasonable price comes from not paying any artists to source the learning datasets. Shady use of 'non-profits' to do that but and hope to avoid legal trouble.

100% agree. I’m just talking about what they actually paid, not what they should be paying.

If artists didn't want their art to be used (for a variety of purposes!) they shouldn't have made it freely available on the internet.

Maybe art websites will add checkboxes to the effect of "don't use this art to train AI models" to submission forms. That won't be enforceable in court (unless the law changes), and it will only deter completely scrupulous actors much like robots.txt.

Art websites might start using javascript to compose images in the browser so they can't be easily downloaded, but that will be quickly defeated and ultimately that arms race is unwinnable.

>If artists didn't want their art to be used (for a variety of purposes!) they shouldn't have made it freely available on the internet.

that isn't how copyright works. I can pirate any book in the world but if I try to pass it or parts of it as my own (minus fair use and time wise) I can be sued for copyright violation. how this applies to AI models is an open question, but "I can freely get it therefore it's ok" is a brain dead legal take

>Maybe art websites will add checkboxes to the effect of "don't use this art to train AI models" to submission forms. That won't be enforceable in court (unless the law changes), and it will only deter completely scrupulous actors much like robots.txt. Art websites might start using javascript to compose images in the browser so they can't be easily downloaded, but that will be quickly defeated and ultimately that arms race is unwinnable.

"it's inevitable" also won't hold up in any sane court

Use != distribution. This isn't a copyright issue. This is a "licensing" issue, in the sense that authors believe they have "licensed" people to view their works, but not to train models on them. That view is incorrect given current law, though there is a small chance the law gets updated to make that position more tenable in court.

I think it's more like authors have not "licensed" people to use their works to make profit from, which is current law. The AI companies just use a 'non-profit' entity to collect the images to skirt around the law, otherwise why would they use a non-profit subsidiary and be very careful about only using non-copyrighted music for music based AI?

I can grind up CDs and use the glittery material to make an outfit that I sell to a fashion collector for a million dollars. Use doesn't matter, transmission of a copyrighted work or something sufficiently derivative is what matters.

The image collection has been primarily performed by Laion, which made up of researchers from various academic institutions. The datasets are very popular in research, and they've enabled a lot of advances in the field.

Copyright law doesn't care about "use". It cares about copying and redistribution. The code at issue here was copied by GitHub and OpenAI for training. A fair use defense is warranted here because if it's successful no license will need to be abided by.

>If artists didn't want their art to be used (for a variety of purposes!) they shouldn't have made it freely available on the internet.

There’s all sorts of software and source code freely available online, but you can’t do whatever you want with it because of the license agreement it’s released under?

How is this any different from software license agreements, or terms of service agreements on websites? People have to check boxes all the time saying they will or won’t do certain things. And those terms are definitely enforceable (in principle).

You are correct that this will eventually be a licensing issue, with art websites being updated to try to prevent people from downloading with the intent to train models. That approach will probably fail as EULAs have been beaten down in courts repeatedly. There's also a loophole there, in that if you downloaded the images from a pirate site, you aren't bound by the EULA, and if you didn't upload data to others you aren't committing copyright infringement.

We have a deep answer and a business answer.

Business answer: Automation saves money; makes humans do more things. The AI-generated content doesn't has to be perfect to be useful. Like 80% good is enough!

Deep answer: We believe artists nor human art isn't going anywhere! We enjoy the process of art being created. Look at Insta/TikTok/Shorts how artists are still growing by sharing their creation process.

Conclusion:

Art for business will be automated. As we said, it only has to fail 20% of the time to be acceptable.

Artists can leverage their personal story and authenticity to still thrive in the era of AI-generated content.

Who believes artists and human art isn't going anywhere? So this business art is just replacing art no one was doing, or it was just happening for free somehow?

Do you think any of the artists that created the content the AI learns from are going to get any royalties for this business use?

This is a bit of a red herring. Nobody in their right mind will just copy another artist 1:1 and make money with it. That would be plagiarism, and that won't have value anyway.

Even imitators (check out how much of those the artists like Ilya Kuvshinov have on Artstation, for example) vary their style a bit. That doesn't save them from being samey and boring, though. Some imitators like those of Jacek Yerka are less boring, yet they still have no message to tell.

Artists have been driving themselves out of the industry for decades, without any ML. What can be easily automated with the current tools (which is not a lot) isn't worth saving at all. I suspect that most people perpetuating the panic never actually dived into AI art, learning what the tools are and aren't good for; they mostly enter a prompt, think it's magic, and start doomposting on Twitter. This is probably how artists felt with early 3D CGI, except we didn't have Twitter back then.

No, I think "it's going to happen anyway because things happened in the past" is the poor argument.

These AI companies are wholesale scraping art without permission, payment or credit. You just have to look at how they act in other, more litigation heavy industries to see what they're getting away with.

I'm not sure how your argument relates to businesses not paying for artists?

>These AI companies are wholesale scraping art without permission, payment or credit.

I specifically mentioned Ilya Kuvshinov, who is one of the most copied artists on Artstation; you can see so much of his rendering, poses, lighting, ideas around that it gets uncanny. Have all those imitators (who make money with it) paid anything to him for scraping his art into their minds? Maybe he got credited? Maybe they asked him for permission? I think it's neither of those.

Artists have been copying each other much more blatantly for ages, and they never had any qualms about it, they've just been diminishing their own value. What's changed?

Artists most certainly do care once they have the finances and value to fight/protect it. Not all of them of course, but there's a difference between heavy inspiration and plagiarism, as per the music industry.

I'd say the key difference with AI is its scale, which can widely out-produce any genuine artist by a massive amount x by many, many multiples of 'AI artist' personas used.

> Like, why are we wasting resources to fix something that human beings actually enjoy doing themselves?

There's entertainment (99% of the stuff) and there's high art. For creators, entertainment is largely driven by money. As a result, especially since 70-80's and artists like Frazetta and Vallejo, most of the entertainment converged into simplified, easily repeatable styles that are easier (=cheaper) to produce, at the cost of quality and thought put into it. However, for other people the interest in entertainment is driven by novelty; same stuff gets boring very fast. This is a fundamental conflict of interests.

With the new tools, it becomes possible to automate the mechanical work like rendering; make a sketch and it turns it into a fully rendered picture in your own visual language that you constructed beforehand. This opens a whole new range of styles that were impractical to use before, due to the amount of work involved. For example, you can make animation in practically any style, not being limited to the striking "3D CGI" look or the simplified cartoonish look, or needing a huge ass team of VFX specialists - the tool does most of the rendering for you without putting specific limits on your style. Long form indie animation becomes possible.

>Am I the only one who finds the current crop of AI startups a little misguided in their priorities?

They are misguided, but probably not in a sense you imply. First, there's plenty of successful ML-based tools in the visual industries. ML-driven keying is pure magic, it's completely accepted in video editing now, it saves huge amount of work and enables things that weren't possible before. Same for denoising, and many other things.

Image generation tools are a bit a solution in the search for a problem, in particular because text to image is a gimmick and is insufficient for anything, and those startups don't seem to realize that. The novelty of prompt-based generation will wear off quickly.

What is needed is style transfer: construct you own visual language, teach it to the model and let the machine use it on a massive scale, leaving you to higher level concepts. This is a product that can be used in production.

Take a look at NVIDIA Canvas, which was a good early demonstration of the principle:

https://www.youtube.com/watch?v=mlZYRwJ2oJg

NVIDIA Canvas was a flop because the model was so basic, it couldn't make anything useful, and most importantly you couldn't teach it new stuff. The current generation of diffusers (Stable Diffusion in particular) is better but isn't quite there yet. You can train the model on your visual style and use it to automate most of the routine stuff, but it still makes mistakes that don't pass the quality test. The current generation isn't optimized for style transfer, the models are too small to be coherent, lack temporal stability for animation etc. Still, open models are very good at honing out the concept and identifying what is really needed.

Reply on news.ycombinator.com