[orbital-decay]

That's not the expectation at all; a lot of work is being done to make it run on underpowered hardware. SD in particular runs on a 8-years-old potato, albeit slowly and with limitations, despite originally barely fitting into 10GB VRAM.

>A theoretical nice thing about Krita and art in these past decades was that you could be an 18 year old with some ok drawing skills, a thinkpad, a secondhand wacom tablet and a version of krita

You never needed a computer for that, just a pen/pencil and paper.

For digital painting in particular though, that only became possible in the recent years. Free digital painting software sucked until recently, so 20 years ago every 18 years old just pirated commercial software. And drawing tablets only became cheap and good after Wacom battery-less patents expired (alternatively, with the advent of iPads with pens that a lot of parents bought for their kids, and cheap drawing software in the App Store).

I'm not even starting on 3D, which always required beefy hardware. Tinkering with Maya/3DSMax/Lightwave in early 2000s required a really powerful gaming PC. These days you can at least rent a powerful GPU for peanuts to run the AI model.

[Fraterkes]

Sure, the part about everyone just pirating photoshop is absolutely tue (it comes out to be the same thing though, you can't pirate hardware). My point is the gap in potential quality and art output between photoshop on a powerful pc an a pirated copy of ps on a thinkpad is pretty small: you need a lot of ram to produce 4k art, but a thinkpad is fine for most comissions. The gap is obviously a lot larger with ai: you yourself mention that sd (just one of the models people are currently using) runs slowly and with limitations. If the expectation becomes that you deliver 100 4k permutations on a certain theme, the time it takes to achieve that from a human labor standpoint will be similar, but the time that takes to render wise will vary orders of magnitude based on your resources. Not to mention that a workflow with a realtime refresh rate is qualititavely different frome one that runs 0.1fps.

[orbital-decay]

Commissions are professional work. If you need speed to get things done, you can pay to rent the GPU and the cost will be negligible to what you earn, even if you're just starting. This is really not that much of a barrier compared to the hoops hobbyist 3D artists had to jump through 20 years ago.

Regardless, you can run SD on a several years old laptop just fine, this is entirely within the reach of most; yes you won't be getting realtime updates but that's not really necessary.

And that's only the beginning. SD was trained using really poor data; everybody is doing that on semi-synthetic datasets with much higher quality labeling now; high quality data and new advancements (see the Beyond U paper [1], for example) allow fitting more into several times less weights with much faster inference. In a year or two, this will be available to practically everyone.

[1] https://arxiv.org/abs/2310.20092