site banner

Friday Fun Thread for December 8, 2023

Be advised: this thread is not for serious in-depth discussion of weighty topics (we have a link for that), this thread is not for anything Culture War related. This thread is for Fun. You got jokes? Share 'em. You got silly questions? Ask 'em.

2
Jump in the discussion.

No email address required.

So you're using them enough that you have to "constantly" correct historical ethnic choices, but also calling them "awful?" If an artist was awful, I would stop asking them for new art and go to someone else. Is your job forcing you to use Dall-e 3 specifically or something? What kind of image/material are you using it for?

I don't have a subscription to GPT 4, so am unable to test this, but the previous iteration allowed users to mention styles that they want it to emulate, and there isn't necessarily an advantage to just leaving it at its default style. If it's oversaturated, you can probably request a limited palette? I tried asking Dall e 2 for a painting with a zorn palette, and it used too much blue (zorn replaces blue with black as a primary), but maybe GPT could help interpret that kind of thing (or I could try spelling out what I mean more clearly?).

I had heard that people have been making add-ons for Stable Diffusion that point it toward specific styles, so that might be worth looking into as well.

Can't beat the cost and convenience for a good-enough image!

I'm sure we'll get to a spot in 2 or 3 years where this gets a lot better. I do have Stable Diffusion but it's slow and hard to wrestle with. I do use Dall-E 3 for work but it's not a large part of what I do. Let's say I generate 2-3 images a day.

My whine here is specifically about the stylistical awfulness of Dall-E 3 images which I now see cropping up everywhere. Prompt-hacking doesn't work. I try stuff like this: "Simple, not complex, no extra characters, restrained, not saturated", but it doesn't seem to really give me what I want.

which I now see cropping up everywhere

I haven't noticed it -- do you have an example?

"Simple, not complex, no extra characters, restrained, not saturated"

Maybe it has trouble with negatives? I wonder if it would respond to directions about specific color palettes (yellow ochre, Paynes grey, cadmium red?), where to place the focal point, or name dropping Rembrandt?

I haven't noticed it -- do you have an example?

Sure. Here are a some examples from a blog that was posted to the slatestarcodex subreddit.

1, 2, 3

Once you recognize the "style" you see it everywhere. The main thing is that they are just way too busy.

I invite you to show me anything that makes all these images I've generated samey.

You're prompting it wrong.

Common link is they are all have far too many unnecessary elements that detract from the image. I will grant that only image #2 looks like a 100% match for the Dall-E 3 archetype.

What do you mean by "unnecessary things"?

They're precisely what I asked for, within the limits of my prompting and the model. Without knowing the prompts, I have no idea what you think they're missing.

At the very least the last one is a minimal brutalist logo for a PMC, I can hardly imagine what could be less so.

While I like (and sometimes exploit!) this trait, a lot of settings on both generation and upscaling (especially with latent upscalers) will result in visual clutter that a normal artist would not use.

This is most noticable and obvious on the PMC brutalist logo: the scattered white pixels around the 'shoulder' and well outside of the logo's boundaries are just not what you'd expect to see. Maybe as some sort of deep-fried jpg artifact, were the rest of the image busier? But they're not actually those things, or even human interpretations of those things.

The wave-face image is the one where clear errors are most human-like -- anatomy and cloth flow mistakes, overpronounced foreshortening, slightly jank perspective are all totally things even good artists do, sometimes intentionally! -- but separately it's also got some weird distractions. Why are there blue highlights on his abs? If the flow of the image is supposed to be toward his face, why are so many lines going to his shoulders?

The ARMA one is the closest to human-like (there's a few physics/layout errors, but they're absolutely ones humans would make), though the genre it's coming from tends to be cluttered and intentionally disorienting to start with.

You can work around and stop these sort of issues, but you have to really heavily ride and push it toward specific low-clutter styles, and even then it takes some futzing with SD parameters to avoid the image coming out overdone or undercooked.

/images/17023968095808215.webp is prompted by meta at the FurryDiffusion discord, but outside of the hands/paws (and... subject matter), it's as close to human-created art as you'll get.