Hacker News Reader

Comment by keepquestioning | Hacker News Reader

ELI5 - why has there been a cavalcade of Stable Diffusion spam on HN recently? What does it all mean?

First, it was recently released, so there’s novelty. Second, the code and model weights were also released so it is open and extensible, which this community loves. Thirdly, these high quality image generation models are mind blowing to most and it’s not hard to imagine how transformative it will be to the arts and design space.

If if has any greater meaning, we might all be a little nervous that it’ll come for our jobs next, or some piece of them. First it came for the logo designers, but I was not a logo designer, and so on.

the other issue is DallE and them Google/Imagen made a big deal about _not_ giving broad access to it and making it an approval-only beta (DallE) or simply you-can't-have-it (imagen).

So people were hyped up.

The stable diffusion model just got officially released recently, and in the last week a lot of easy to install repositories have been forked off the main one, so it's very accessible for people to do this at home. Additionally, the model is very impressive and it's a lot of fun to use it.

How does it compare to DALL-E

Worse at image cohesion and prompt matching, but competitive in terms of final image quality in the better cases.

Its image quality is often better, mainly because you can run it on your own machine and increase the quality/time controls. DALLE2 lacks settings and doesn't run enough diffusion steps to have fine detail.

Having had access to DallE for awhile, I find both Midjourney and stable diffusion to be quite a bit more powerful.

For weird inputs I find DALL-E to be better, but both have failed at really specific stuff that I've tried to create as art. ie: a factory with human lungs.

With stable diffusion, it really creates some nice stuff for what is already available, like a pizza with specific toppings [0]. I've been using it to add pictures to any of the recipes that are added to my wiki site without a picture. I originally tried this with DALL-E with similar prompts and the results are less appetizing.

[0] - https://www.reciped.io/recipes/mushroom-and-onion-pizza/

Some arbitrary comparisons I did:

Piano keyboard: https://mobile.twitter.com/skybrian/status/15629121632697303...

Model train layout: https://mobile.twitter.com/skybrian/status/15629200301441064...

Turtle sand sculpture: https://mobile.twitter.com/skybrian/status/15640856320323174...

It's an impressive new technology, and there's nothing else out there like it in terms of the model being publicly available and able to be run on consumer GPUs.

You can now fire artists/designers and replace them with AI. Obviously, that's cheaper.

Someone will surely come by soon and tell us, “well actually… artists and graphic designers are irreplaceable.”

But for real, plenty of people are going to start rolling their own art and skipping the artist. Not Coca-Cola, but small to medium businesses doing a brochure or PowerPoint? Sure!

I think there's going to be plenty of work in stacking multiple AI prompts or manual retouching to fix rough spots. It automates a task, not a job. Some people won't use it at all and other people will use it only for reference - in the end doing everything by hand, as usual, because they have more control and because AI art has a specific smell to it and people will associate it with cheap.

But it's not just for art and design, it has uses in brainstorming, planning, and just to visualise your ideas and extend your imagination. It's a bicycle for the mind. People will eat it up, old copyrights and jobs be damned. It's a cyborg moment when we extend our minds with AI and it feels great. By the end of the decade we'll have mature models for all modalities. We'll extend our minds in many ways, and applications will be countless. There's going to be a lot of work created around it.

>AI art has a specific smell to it and people will associate it with cheap.

It might now, but I feel like that will be trained out of it a few more papers down the line

People rarely write assembly code nowadays because we mostly all use higher level abstractions that let us write more powerful code with fewer lines.

There are plenty of small shops now where somebody knows a little Photoshop and can eek out a design that they otherwise wouldn't be able to using pen and paper.

There are also professionals that use the Adobe suite to enhance their abilities they've cultivated for years.

AI art will simply be a tool that enhances artists but might take away some low hanging fruit jobs similar to how web frameworks pushed people out of the job of webmaster and into more specific roles.

If you automate part of a job, you need fewer people doing the job. We still have farmers, but we automated enough of the job to do the same work with a thousand times fewer farmers.

I think the first business to crack will be stock image sites like shutterstock.

Those were already used in a "let's find something that roughly fits what I want to communicate with this text" way.

Today I created a quick get well soon card using an image from the new Midjourney beta and I have to say the result was exactly as good as if I had used Shutterstock but it took me much less time because the search prompt matched created something I wanted on the third try.

Comparing that to sifting through pages and pages of vaguely relevant images it's a clear win and a lot cheaper

For sure it automates some work. For example, my sometime hobby of making silly photoshops looks like it will now be a whole lot easier... Visual memes can just be a sentence now. For more serious work I wonder... But it does give pause about what it means for other forms of work.

It'll be an interesting line to be sure.

Right now the tech still requires some nuance to be able to slap it all together into what I think most people would want.

While i expect the interface and the like to get a lot better, all good tutorials of this tech so far show many iterations over many different parts of an image to get something "cohesive". Blending those little mini iterations together is VASTLY easier than just making the whole thing, but not just plug and play for something professional.

Still there will be a huge dent in how long it takes to make certain styles of work and that will lower demand considerably, and there's a large market of artists who thrive on casual commissions which this might replace.

Weren’t the small to medium businesses already mostly using stock images anyway?

If anyone was to be worried I’d think it would be Getty Images.

There is plenty of reason for artists who are hired for one-of-a-kind work to be worried.

Getty Images will just get with the program and stock up on bazillions of AI generated images, indexed by the prompts used to generate them.

Someone looking for stock images doesn't want to deal with artists, photographers or feeding prompts after prompt into some AI software, while not quite getting the desired result.

If Getty makes it easier for someone to find some existing AI-generated image than to generate one, they still have something.

A lot of the AI images we see in online blogs and galeries have been curated; people tinkered for hours with the stuff, and cherry-picked the best results. There could be some business model in that, at least for a while.

Getty would just be a cache with the prompts acting as the initial search index and buyers are just typing keywords in to buy cached images.

That is not exactly so, because the prompts are not reproducible input cases; you don't get the same image every time for a given prompt. An association between a prompt and, say, around ten images would be something resembling a cache.

Reply on news.ycombinator.com