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

No elephants: Breakthroughs in image generation

oneusefulthing.org

Post reply on HN
Print view View original

31–40 of 373 posts

Re: No elephants: Breakthroughs in image generation

#31

Post by GaggiX »

card_zero wrote: post #10
x187463 wrote: post #3

This is a before/after moment for image generation. A simple example is the background images on a ton of (mediocre) music youtube channels. They almost all use AI generated images that are full of nonsense the closer you look. Jazz channels will feature coffee shops with garbled text on the menu and furniture blending together. I bet all of that disappears over the next few months. On another note, and perhaps other…

I was wondering yesterday how AI is coming along for tweening animation frames. I just did a quick search and apparently last year the state of the art was garbage: https://yosefk.com/blog/the-state-of-ai-for-hand-drawn-anima... Maybe this multimodal thing can fix that?

That blog post is a year old.

There has been a lot of progress since then: https://doubiiu.github.io/projects/ToonCrafter/

Re: No elephants: Breakthroughs in image generation

#32

Post by thrance »

Each generation follows the prompt a little bit better than the last, but I don't see any revolutions. Fingers are still messed up, eyes are wonky and legs sometimes still fork into two. Fundamentally it's still the same diffusion technique, with the same limitations.

Re: No elephants: Breakthroughs in image generation

#33

Post by Taek »

probably_wrong wrote: post #12

> Is it okay to reproduce the hard-won style of other artists using AI? Who owns the resulting art? Who profits from it? Which artists are in the training data for AI, and what is the legal and ethical status of using copyrighted work for training? These were important questions before multimodal AI, but now developing answers to them is increasingly urgent. I have to disagree with the conclusion. This was an importa…

I don't think there's consensus around that idea. Lots of people (myself included) feel that copyright is already vastly overreaching, and that AI represents forward progress for the proliferation of art in society (its crap today, but digital cameras were crap in 2007 and look where they are now).

Its also not clear for example that Studio Ghibli lost by having their art style plastered all over the internet. I went home and watched a Ghibli film that week, as I'm sure many others did as well. Their revenue is probably up quite a bit right now?

"How can we monetize art" remains an open question for society, but I certainly don't think that AI without restrictions is going to lead to fewer people with art jobs.

Re: No elephants: Breakthroughs in image generation

#34

Post by gcanyon »

It's interesting to hear people side with the artists when in previous discussions on this forum I've gotten significant approval/agreement arguing that copyright is far too long.

As I've argued in the past, I think copyright should last maybe five years: in this modern era, monetizing your work doesn't (usually) have to take more than a short time. I'd happily concede to some sort of renewal process to extend that period, especially if some monetization method is in process. Or some sort of mechanical rights process to replace the "public domain" phase early on. Or something -- I haven't thought about it that deeply.

So thinking about that in this process: everyone is "ghiblifying" things. Studio Ghibli has been around for very nearly 40 years, and their "style" was well established over 35 years ago. To me, that (should) make(s) it fair game.

The underlying assumption, I think, is that all the "starving" artists are being ripped off, but are they? Let's consider the numbers -- there are a handful of large-scale artists whose work is obviously replicable: Ghibli, the Simpsons, Pixar, etc. None of them is going hungry because a machine model can render a prom pic in their style. Then you get the other 99.999% of artists, all of whose work went into the model. They will be hurt, but not specifically because their style has been ingested and people want to replicate their style.

Rather, they will be hurt because no one knows their style, nor cares about it; people just want to be able to say e.g. "Make a charcoal illustration of me in this photo, but make me sitting on a horse in the mountains."

It's very much like the arguments about piracy in the past: 99.99% of people were never going to pay an artist to create that charcoal sketch. The 0.01% who might are arguably causing harm to the artist(s) by not using them to create that thing, but the rest were never going to pay for it in the first place.

All to say it's complicated, and obviously things are changing dramatically, but it's difficult to make the argument that "artists need to be compensated for their work being used to train the model" without both a reasonable plan for how that might be done, and a better-supported argument for why.

Re: No elephants: Breakthroughs in image generation

#35

Post by hnbad »

probably_wrong wrote: post #12

> Is it okay to reproduce the hard-won style of other artists using AI? Who owns the resulting art? Who profits from it? Which artists are in the training data for AI, and what is the legal and ethical status of using copyrighted work for training? These were important questions before multimodal AI, but now developing answers to them is increasingly urgent. I have to disagree with the conclusion. This was an importa…

> What the post should say is "we know that this is unfair to artists, but the tech companies are making too much money from them and we have no way to force them to change".

It seemed a fact of life that companies will just abuse your personal data to their liking and can do what they want with information they collect about you because "if it's free, you're the product" (and even if you paid for it, "you should know better" etc). Then GDPR and its international derivatives came along and changed that.

It seemd a fact of life that companies that technically don't have an actual market monopoly can do whatever they want within their vertically integrated walled gardens because competitors can just create their own vertically integrated walled gardens to compete with them and the rules for markets don't apply to walled gardens. Then the DSA and DMA came along and changed that.

I don't see why legislation can't change this, too. Of course just with the GDPR, DSA and DMA we'll hear from libertarians, megacorps and astroturf movements how unfair it all is to mom & pop startups and how it's going to ruin the economy but I think given the angle grider the US is currently taking to its own economy (and by extension the global economy because we're all connected), I think that's no longer a valid argument in politics.

Re: No elephants: Breakthroughs in image generation

#36

Post by qiqitori »

Wha- wha- what? I tried to generate an image in ChatGPT after the announcement a while back and the image wasn't bad, but the text on it (numbers) was nonsense. (Analog gauge with nonsense numbers instead of e.g. 10, 20, 30, 40, etc.)

Gave it another chance now, explicitly calling out the numbers. Well, they are improved but not sure how useful this result is (the spacing between numbers is a little off and there's still some curious counting going on. Maybe it kind of looks like the numbers are pasted in after the fact?

https://chatgpt.com/share/67f4fa33-70dc-8012-8e1e-2dea563d3d...

Re: No elephants: Breakthroughs in image generation

#37

Post by danielbln »

rwmj wrote: post #18
fxtentacle wrote: post #9

Earlier quoted context omitted.

Imagine you cut the image into 32x32 pixel blocks. And then for each block, you can chose 1 out of 128,000 variations. And then a post-processing step smoothes out the borders between blocks and adjusts small details. That's basically how a transformer image generation model works. As such, the process is remarkably similar to old fixed-font ASCII art. It's just that modern AIs have a larger alphabet and, thus, more…

I don't get how this would produce consistent images. In the article, the text could be on a grid, but the window and doorway and sofa don't seem to be grid-aligned. (Or maybe the text is overlaid?)

The model looks ahead, just like LLMs look ahead. An LLM outputs token by token but can still output a fully coherent and consistent story for example. This new crop of auto-regressive image models does the same.

Re: No elephants: Breakthroughs in image generation

#38

Post by Taek »

gcanyon wrote: post #34

It's interesting to hear people side with the artists when in previous discussions on this forum I've gotten significant approval/agreement arguing that copyright is far too long. As I've argued in the past, I think copyright should last maybe five years: in this modern era, monetizing your work doesn't (usually) have to take more than a short time. I'd happily concede to some sort of renewal process to extend that p…

I find it unlikely that someone who was willing to pay an artist for a charcoal sketch would be satisfied with an AI alternative.

You don't just buy art for the aethstetic, you buy it for a lot of reasons and AI doesn't give any of the same satisfaction.

Re: No elephants: Breakthroughs in image generation

#39

Post by hnbad »

Taek wrote: post #33
probably_wrong wrote: post #12

> Is it okay to reproduce the hard-won style of other artists using AI? Who owns the resulting art? Who profits from it? Which artists are in the training data for AI, and what is the legal and ethical status of using copyrighted work for training? These were important questions before multimodal AI, but now developing answers to them is increasingly urgent. I have to disagree with the conclusion. This was an importa…

I don't think there's consensus around that idea. Lots of people (myself included) feel that copyright is already vastly overreaching, and that AI represents forward progress for the proliferation of art in society (its crap today, but digital cameras were crap in 2007 and look where they are now). Its also not clear for example that Studio Ghibli lost by having their art style plastered all over the internet. I went…

Copyright is a logical consequence of property rights. I'd agree that property rights hold back industry and trade but if you want to abolish property rights, you first have to decommodify the essentials like food, housing, public infrastructure and healthcare, because unleashing the market when it has control over all of these is going to have some very undesirable consequences.

Re: No elephants: Breakthroughs in image generation

#40

Post by hnbad »

YurgenJurgensen wrote: post #11
roenxi wrote: post #2

That proper "no elephants" first image is hilarious. Another key point here is the generative AI's meme game is getting rather strong. Which isn't a small thing, humour is an advanced soft skill.

Having thousands of copies of that image in your training set isn’t a skill at all.

AI's "meme game is going strong" by the same metric I would use to try to argue that Elon Musk's is.

I wouldn't call it a good metric, though.

Post reply on HN