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

AI doppelgänger experiment – Part 1: The training

julienposture.substack.com

Post reply on HN
Print view View original

101–110 of 137 posts

Re: AI doppelgänger experiment – Part 1: The training

#101

Post by Der_Einzige »

GaggiX wrote: post #53
mattkevan wrote: post #47

Earlier quoted context omitted.

As a follow on: One of my friends is a professional illustrator and was concerned about the impact of AI on their work. As a demonstration I trained some models on their styles. They were worried right up until they saw the output. Yes, it looked superficially like their work, but the slightest glance showed it to be absolute crap. Anyone who would consider using the AI version of their work over commissioning the ac…

It's possible that your model wasn't great, for my experience, in the past I uploaded an AI image in the same style of a certain artist to a website that organizes anime images, someone tagged it with the artist's tag because they thought it was made by him, also the image I uploaded got much more "likes" than anything made by the artist in months, and it's not like I'm spamming images, that was the only upload. (I l…

Just call them what they are, "booru" websites. The AI community needs to actually understand what these are and why they just happen to give us extremely large, high quality, well taged datasets. Yes, the AI community will have to admit that coomers have been really good for AI development. Yes that will rustle feathers.

A lot of AI researchers are either playing dumb or are actually ignorant of this space. A lot of really talented folks are operating in the shadows and should be at NeurIPS and ACL presenting and bartering for VC funding rather than posting anime waifus on /g/ and staying anonymous on discord...

Re: AI doppelgänger experiment – Part 1: The training

#102

Post by lz400 »

vunderba wrote: post #18
voiceblue wrote: post #4

Related: "Invasive Diffusion: How one unwilling illustrator found herself turned into an AI model" [0]. It's astonishingly simple to train both LoRAs and Embeddings nowadays -- if you combine it with tech like ControlNet, you can more or less have direct control over the illustration as well (see this example I converted to Hollie Megert's style with barely any effort [1] using the model from [0]) -- but FWIW I was p…

Corporate Memphis seems to have that very "flat almost clip-art" style to it. Generative art is ideal for this type of stock art use case where 90% is simply good enough for the vast majority of situations. I feel like the gig economy (things like Fiverr) are unfortunately going to get absolutely crushed by generative art systems since most of their customers aren't that particular and are likely fine with anything t…

This comment thread could be from 3 years ago. What you're predicting I think has for the most part already happened. Everyone I know who used to commission Fiverr stuff has switched to AI.

Re: AI doppelgänger experiment – Part 1: The training

#103

Post by leobg »

gedy wrote: post #7

The challenge I see is that (in the US at least), "style" is not copyrightable in the arts. IMHO this is a good thing because artists learn and copy from one another, and if it were copyrightable that would lead to no end of lawsuits between media companies. So my concern is trying to apply copyright to training AI models can easily slippery slope into more draconian copyright rules for human artists.

One reason for that is that, like with ideas, society benefits, from styles intermixing. So making them "copyrightable" would be to the detriment of society.

But the other reason might be that it has just not been feasible to actually measure things like “style“. And if you cannot, in your lawsuit, put into words what you claim is, and where its boundaries are, it’s impossible to enforce it.

It’s very much possible that people will forget the former reason and jump on the latter, now that one might be able to use AI to actually quantify, measure, and map things like style. In fact, I wouldn’t be surprised to see companies like OpenAI play Good Samaritan and “helping legislators“ solve the problems generated by AI. Only after they have secured the own datasets, of course.

Re: AI doppelgänger experiment – Part 1: The training

#104

Post by Eisenstein »

ecjhdnc2025 wrote: post #40
jsheard wrote: post #8

Earlier quoted context omitted.

> because as soon as there's enough work in the new style it can be stolen. This is where the oft-cited comparison to previous advances like the camera breaks down - AI has an ongoing parasitic relationship with the mediums it is displacing. The camera doesn't need a thousand photorealistic paintings of a new subject to be made before it can capture that subject going forward. AI is an unprecedented situation, as far…

I agree wholeheartedly. As a photographer I am shocked to see other photographers who are AI proponents making the comparison to the beginning of the photography era. (Though it's usually the very bad photographers or those who haven't actually done any reading about the impact of photography on art who make this point) Art's one hope to avoid this parasitism is to move back into the realms of the purely physical, I…

The fact is that it is much like that, because no one then knew what was going to happen and it seemed to them like an existential threat. The mechanism of how it works or its relationship with our previous technology is irrelevant. The comparison is about everyone freaking out because no one knows how this is going to play out.

But history has shown time and again that carving out a niche in art by using technology is always fleeting and any security you think you had was illusory. Just looking at Hollywood, you have: (a) transition to sound (b) emergence of televisions (c) transition to color (d) effects revolution (e) digital replacing practical effects (f) digital replacing film

Each one of those transitions obsoleted a generation of specialists with niches, yet Hollywood and the people it employs are still doing just fine.

Re: AI doppelgänger experiment – Part 1: The training

#105

Post by optimalsolver »

astrange wrote: post #90
monadINtop wrote: post #81

Earlier quoted context omitted.

No, you need intelligent things to be intelligent. And it would probably be good to distinguish between "things that appear superficially intelligent" and "things that have genuine capacity for internal experience" while you're at it, since the popular perception on this forum seems to be to just declare that physicalism/materialism is obviously correct and explanatory and we can just brush the hard problem of concio…

Chomsky's critique of LLMs was somehow not even correct about his own theories ( https://x.com/profraha/status/1634311282135318529 ). And of course has the usual problem of 1. claiming LLMs are random statistical models (they aren't, the sampler is) 2. claiming instead of generating correct text they only "seemingly" generate correct text and not explaining why this is any less impossible. But people have built expli…

I swear there's a personality type that simply can't resist injecting their politics into unrelated conversations.

It's truly fascinating.

Re: AI doppelgänger experiment – Part 1: The training

#106

Post by grob-gambit »

Der_Einzige wrote: post #100
voiceblue wrote: post #4

Related: "Invasive Diffusion: How one unwilling illustrator found herself turned into an AI model" [0]. It's astonishingly simple to train both LoRAs and Embeddings nowadays -- if you combine it with tech like ControlNet, you can more or less have direct control over the illustration as well (see this example I converted to Hollie Megert's style with barely any effort [1] using the model from [0]) -- but FWIW I was p…

Can we just talk about the fact that the most popular fine-tuned model for SDXL is called PonyDiffusion ? The circumstances of its creation, and the whole entire world of like, booru websites that you have to explain to boomers in suits with VP titles who are trying to invest in this space is just surreal. It's like a reverse of when they were raising their kids and giving us "the talk" as children. Now we have to gi…

[flagged]

Re: AI doppelgänger experiment – Part 1: The training

#107

Post by birracerveza »

mattkevan wrote: post #41

I’ve been using generative AI in my art for a good few years now and have a few thoughts on this. Yes it’s trivial to train a model on a particular style, I’ve done it many times, including on my own art. It can make things that look pretty good! But it can’t make anything great, not without a lot of luck or creative input from the user. It can make something that looks similar to a style, but it can’t explain what m…

>It won’t kill creativity, it’s just another creative tool.

I've been banging on this drum since AI's inception, and will continue do so.

Even if machines were capable of directly reading our mind and outputting a perfect representation of exactly what we wanted at the mere thought of it, it's still a mere tool bound by the creativity of its artist, even if the artist were the AI itself.

Art is dead. Art is always dead.

Did people complain when digital paintings became a thing?

Re: AI doppelgänger experiment – Part 1: The training

#108

Post by Zambyte »

mattkevan wrote: post #41

I’ve been using generative AI in my art for a good few years now and have a few thoughts on this. Yes it’s trivial to train a model on a particular style, I’ve done it many times, including on my own art. It can make things that look pretty good! But it can’t make anything great, not without a lot of luck or creative input from the user. It can make something that looks similar to a style, but it can’t explain what m…

> But it can’t make anything great, not without a lot of luck or creative input from the user.

That seems true of "non-generative" art too, no?

Re: AI doppelgänger experiment – Part 1: The training

#109

Post by Zambyte »

jsheard wrote: post #8
ecjhdnc2025 wrote: post #5

Earlier quoted context omitted.

> The art masters cannot be replaced, you cannot predict what they’re going to do next because they change the rules entirely. This is how we know a great artist from a mediocre artist. I agree that in principle the best working artists are those who know how to create, modulate and discard artistic languages over time -- creating and abandoning rules for sequences of work -- but AI systems can clearly be unethically…

> because as soon as there's enough work in the new style it can be stolen. This is where the oft-cited comparison to previous advances like the camera breaks down - AI has an ongoing parasitic relationship with the mediums it is displacing. The camera doesn't need a thousand photorealistic paintings of a new subject to be made before it can capture that subject going forward. AI is an unprecedented situation, as far…

Do we want art to be highly valued? Doesn't that make it out of reach for the masses? I think making art more approachable is a good thing. I probably would not have gotten into visual art if I could not use a camera, and had to paint or even dance. The camera de-valued painting by saturating the visual art ecosystem. Image generation does the same once again.

Consider: generative art technologies (as all good technologies) raise the skill floor, without lowering the skill ceiling.

Re: AI doppelgänger experiment – Part 1: The training

#110

Post by mistermann »

ninetyninenine wrote: post #62
mistermann wrote: post #52

Earlier quoted context omitted.

"We're at the tippity top of the mountain, but we're only halfway up".

We're at the foot of the mountain. We only just started.

Surveys of humans strongly demonstrate that they are rather impressed with their mountain climbing accomplishments thus far. The main shortcoming to humanity is only the actions of those other people.
Post reply on HN