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

Stable Diffusion is a big deal

simonwillison.net

Post reply on HN
Print view View original

321–330 of 488 posts

Re: Stable Diffusion is a big deal

#321

Post by salawat »

jeroenhd wrote: post #299
HelloNurse wrote: post #286

Playing with the prompt-only demo at https://huggingface.co/spaces/stabilityai/stable-diffusion I got the impression that many apparently harmless requests corner the model into a very sparse set of examples exhibiting extreme bias and utter nonsense. For example (seed 0, other advanced options at default values): Blue hamsters filling donuts with nails mostly edible donuts and quasi-donuts; some blue, but no hamster…

I think many of your prompts have trouble getting the right representation because you're using a lot of abstract and implied language. For these prompts to work right, you need to be very specific and often add keywords as if you're specifying a mood board. You also need some overlap between what the model has likely been trained on and the expected output. The system works much better in img2img mode. Slapping toge…

>You also need some overlap between what the model has likely been trained on and the expected output.

Ooooo. Ouch. That's... Kind of a death knell for a good tool in my experience, and suggests the tool in question is a glorified search engine in a sense. With a really confusing query syntax composed of words, and graphical starting states.

If I have to become to get anything done... Why not just learn to draw/hire a creative?

Re: Stable Diffusion is a big deal

#322

Post by mountainriver »

_gabe_ wrote: post #307

Is it just me, or do the comments in this thread seem to be the exact opposite of the sentiment in the comments on similar Github Copilot threads? I just find it a bit ironic that programmers are irate about Github Copilot using their copyrighted material to train. However, if it's an ML model training off of copyrighted artists material, clearly its a transformative work. I just find the opposing sentiments for thes…

Yeah I hadn’t thought of it that way. Curious how designers feel about image generation. I imagine everyone will love AI that doesn’t threaten their job

Re: Stable Diffusion is a big deal

#323

Post by Swizec »

stocknoob wrote: post #92
michaelchisari wrote: post #14

Earlier quoted context omitted.

That’s where it’ll be truly disruptive. SD for ideation, humanity for the final product.

Honestly, I don’t even know if we’ll need humanity for the final product. It’s like using a chess computer to get an idea of a good move… and then a human to approve it? Adjust it? Be inspired by it? Any signals humans give (painting x is better than y) is another signal to encode. Take billions of such ratings and improve the AI’s taste to superhuman levels. In short, anything that humans would add to artistically i…

> It’s like using a chess computer to get an idea of a good move

They did this for Backgammon and found styles of play humanity hasn’t discovered in millennia. Now humans can use those when playing each other and it makes an old boring game feel fresh and exciting.

Re: Stable Diffusion is a big deal

#324

Post by ChrisMarshallNY »

dsign wrote: post #64
Fordec wrote: post #11

Earlier quoted context omitted.

For every tool, a better output will always be created by people who specialize in a craft. Just as photoshop revolutionized photography, to this day you can tell the difference easily between and bad 'shops. In video games upon release everyone is bad and its an even playing field. But as people practice and gain experience they improve their usage, refine their approaches. You eventually see metas develop and best…

I agree, this would be an incredible tool. I can see how some of the outputs may help me improve a piece I'm working on, even if I would never use the model's output for my final product.

It looks like a decent way to produce concept sketches.

Currently, working with creatives (including programmers) is a very iterative process for non-creatives.

"I want this."

"No, I meant this."

"Can we try making that line longer?"

"Eh, I'm not feeling it. Why don't we try brighter colors?"

"Ugh. That looks obnoxious. Can we tone down the red?"

etc.

If the non-creative can use this system to reduce that iteration, it will make life easier for them (maybe not for the creative, if they are banking on the hours doing the iterations).

Re: Stable Diffusion is a big deal

#325

Post by scarmig »

_gabe_ wrote: post #307

Is it just me, or do the comments in this thread seem to be the exact opposite of the sentiment in the comments on similar Github Copilot threads? I just find it a bit ironic that programmers are irate about Github Copilot using their copyrighted material to train. However, if it's an ML model training off of copyrighted artists material, clearly its a transformative work. I just find the opposing sentiments for thes…

I think both are fine. One source of the difference in reception (aside from the selfish aspect) is that the space of work generated by Copilot has fewer degrees of freedom than visual art. So the things it produces are much more likely to be near-identical to existing code than Stable Diffusion is to images. I.e. if SD were just providing near identical copies of artists' work, people would be more sympathetic.

Re: Stable Diffusion is a big deal

#326

Post by peoplefromibiza »

hbn wrote: post #311
peoplefromibiza wrote: post #302

Yeah, it is. It is for authors that got their works stolen for "the greater good" without even being notified. A friend of mine found his works in the Stable Diffusion dataset, the work was not meant for public use, he's never been notified and, most of all, he would have never agreed if they cared to ask. https://laion-aesthetic.datasette.io/laion-aesthetic-6pls/im...

I don't think the word "stolen" applies here, at least not any more-so than I stole from him when I looked at that image. I'm not sure about that website rehosting his image, but in terms of the model people are running on their computers, it's not like it literally includes the image that can be pulled back out of it. A computer looked at that image, and tweaked a very small amount of numbers within ~5GB of numbers…

> not any more-so than I stole from him when I looked at that image

I spoke with the author, the author disagrees with you.

The reason he consider that stealing is because they took that image from the online shop where the disclaimer specifically states "2022 © Magnetic Press LLC by Arenathemes. All Rights Reserved" without even trying to contact the copyright holder.

https://store.magnetic-press.com/products/viewpoint-by-lrnz

> But when I publicly put my code out there, I was also giving people the opportunity to learn from my code and reproduce similar things to it as well.

Not really, IMO

I can enter a bookshop and look around, open a book, read some pages and maybe learn something.

It doesn't give me the right to republish a similar book putting together what I saw, unless the license specifically grants me that right or I obtained a license that does.

Code Pilot is a commercial product, free licenses don't automatically grant the right to commercial use of that code.

Re: Stable Diffusion is a big deal

#327

Post by swalsh »

habosa wrote: post #121

These new models completely changed my mind about how much impact AI will have in my lifetime. They are the most impressive software achievements in decades and anyone who has a “meh” reaction will absolutely end up looking silly. I’m not optimistic that their impact will be positive though.

This is happening with images first, but I can see how this might be able to be leveraged for all kinds of things. The speed at which society might be transformed could be astounding. But it won't be distributed evenly, not everyone is prepared to take full advantage of these benefits. This is early 2000's internet distruption x2 in half the time.

BUT it will likely be disrupting the upper middle class more than the lower middle classes, as it's mostly disrupting to creative work. Upper middle class people have a louder voice.

Re: Stable Diffusion is a big deal

#328

Post by throwawaylinux »

osullip wrote: post #282
fingerlocks wrote: post #278

Earlier quoted context omitted.

You included “cobbler” only because it sounds old-timey to you, and I was pointing out that it’s not. Repairing expensive shoes is not an automated process. It’s more like fixing a roof leak, landscaping, or changing a flat tire. Jobs for those things still exist and aren’t going anywhere.

And people who have open fires often use the services of a chimney sweep. It's how you keep your chimney clean. https://melbournechimneysweeping.com.au/ is one near me. Not wanting to pile on the GP but their point is moot.

I said chimney sweeps still exist, so your point is redundant.

Re: Stable Diffusion is a big deal

#329

Post by trention »

varnaud wrote: post #300
trention wrote: post #183

Earlier quoted context omitted.

My position is perfectly logically consistent. Art is distilled human experience and human emotions emerging in a particular context after a chain of events. None of this is true of AI "art". What is not logically consistent is to claim that a black box utilizing statistical relationships between pixels in a giant dataset is an "artist" and that its products create "value". The compiler is not a programmer, AI can ne…

The chain of events: Step 1: Tweak settings and type text Step 2: Look at the result Step 3: If you like the result, go to step 4, otherwise go back to step 1. Step 4: Save and share the result Feels like art to me. Ultimately it's still a human using a tool to create art. For me AI art is just the name of the art style.

By that logic, a google image search + sharing results you like is also art. But it really isn't. At best it's curation.

Re: Stable Diffusion is a big deal

#330

Post by pclmulqdq »

_gabe_ wrote: post #307

Is it just me, or do the comments in this thread seem to be the exact opposite of the sentiment in the comments on similar Github Copilot threads? I just find it a bit ironic that programmers are irate about Github Copilot using their copyrighted material to train. However, if it's an ML model training off of copyrighted artists material, clearly its a transformative work. I just find the opposing sentiments for thes…

If you use my copyrighted code to train your brain and reproduce exact copies of functions I wrote, I have a problem with that. Same for art. Exact copying is the copyright problem.

If you use my copyrighted (public) code to train your brain, and then produce new functions that do new things, there is no problem.

Replace "brain" with "AI" and you get my position on this (and a lot of others' positions). It seems to be a lot easier to transform art than code.

Post reply on HN