I am talentless and untrained. Now with a combo of prompts and img2img, I can create awesome results on any topic and in any style that I have the rights to use. That’s a 0 to 1 moment. It didn’t get faster for me, it went from impossible to possible.
"Any style" seems like an enormous stretch. There definitely seems to be some styles that AI favors, ones which I've seen described as "clutter the frame so you don't notice the flaws". It struggles with simpler styles. I have yet to see a flat black & white image generated by an AI that looked even passable.
Have you tried DALL-E or Stable Diffusion yet? I bet you could generate a black and white image that met your standards for being impressive, if you spent a few minutes on it.
Nah, that's not it. I mean flat #000000/#ffffff. Google "stencil image" and you'll see what I mean.
AI really doesn't handle styles with restrictions like that well. I tried the stable diffusion website with variations on "black and white silhouette stencil image of a cat". It kept wanting to give the cat colored eyes, or it used shading, or the cat didn't have a coherent anatomy, along with the typical AI art "duds" that aren't really anything at all.
To be fair, I did get a couple of passable results when I replaced 'cat' with 'dog'. They were simple, but didn't have any obvious errors.
To be fair in the other direction, replacing 'cat' with 'abacus' gave me an (admittedly pretty) grid of numbers and some chainmail, and 'helicopter' suggested a novel design where two helicopter bodies would be stacked vertically, connected by a vertical shaft through the rotor, and which turned into a palm tree trunk above the top unit. Once you get out of the sample data, it starts to fall apart.
I feel like other people here are willing to forgive more errors than I am. They see an incoherent splotch in an image and assume more development can iron out all the problems, and I see a unavoidable artifact of the fact that these systems don't have a real understanding of what they're making.
Is this not an acceptable result to you? Did this on my first try and to my eyes it’s the same thing as I see when I googled “stencil image.” I’m thinking you have just not tried these prompts enough.
I gave this another shot to see if I could get a more complex stencil. This was my very first try again, so truly not cherry picked. Prompt was: “Stencil image of a tiger face. Clip art. Vector art.” This looks like an infinite stencil making machine to me
That first one just sucks. I don't have anything else to say about it.
The second one is representative of the upper end of what I was getting. It's almost passable, but doesn't hold past about five seconds. The left and right half don't look like they belong to the same animal. The blank space in the middle of the face is huge and detracts from any sense of structure, and the whole mouth area is just odd.
It's seriously impressive for AI, but it's not end-of-artists type stuff. I can google "tiger black and white stencil" and get a bunch of tiger faces, and every one of them is noticeably better. People imagine there are plenty of art jobs where discrepancies like this don't matter, but there really aren't.
Yes, that tiger is shit. Here are 35 tiger stencils I generated over a few minutes with Stable Diffusion on my PC with a humble RTX 2070s gpu: https://i.imgur.com/y9oCZIz.jpg
I could run pretty much any of these through adobe illustrators auto trace and end up with an amazing vector image.
I could also leave it generating these for an hour and I'd have over 1000 results to choose from.
I do feel like you just moved the goalposts on me from AI can’t produce this style at all, to can’t produce this style on a level with an experienced human artist. I don’t think anyone is claiming it beats a good human artist. That does not make it useless for stencils.
Again those were my first try and I know nothing about stencil beyond what 2 seconds on Google images could give me. Certainly better than I could produce if you gave me Adobe Illustrator and a weekend. And the image is mine to use as I please, unlike what I could rip off Google Images.
Also, I thought the cat was cute, but there’s really no accounting for taste. Here’s a silly and swirly cat that might be more your thing? This was a cherry picked one of 10 since you have high standards;)
I considered qualifying “any” but decided it required no qualification. I don’t know how many examples are needed in the training data for a given artist to be able to reproduce that artist, but given how many obscure artists I have seen Dall-e and Stable Diffusion able to recognize, it must not require that many examples. And it’s still possible to fine tune a model with additional training if a new artist comes along or you want a bit more capability with a rare style.
