D&D General AI Art for D&D: Experiments

[Hussar]

Silly Dragon Cover Homage:

View attachment 355497

Damn, that's pretty scarily spot on. I actually HAVE this issue.

 

[Scribe]

 

[overgeeked]

 

[ilgatto]

Those and the others are sweet. What was your prompt?

It's been a while and I'm afraid I don't remember what prompts, exactly, I used for what images.

I did keep many of the prompts for the marshalling thistles, though, and I ended up with things like:

"A drawing of a group of purple thistles marching like soldiers in a medieval French mountain meadow, sunset, line art in colour, (ligne claire, strong lines, no hatching), by Edgar P. Jacobs, by Jacques Martin"

and:

"A drawing of a regiment of purple thistles marching against a medieval French village in a mountain meadow, sunset, marshalling thistles, line art in colour, (((ligne claire, strong lines, no hatching))), by Edgar P. Jacobs, by Jacques Martin"

These were the result of many earlier attempts and I had to actually start describing acts and scenes that would get the AI to understand the concept of "marshalling thistles", which it just wouldn't.

 

[ilgatto]

Those and the others are sweet. What was your prompt?

Okay, also recovered some of the prompts for the dead trees. At some point I found Neural.Love to be capable of generating quite pleasing images of trees for some reason, so I started trying to make images of AD&D monstrous trees.
In this instance, I was aiming for the (Athasian) psionic tree and it seems that I'd been able to get it to generate line art images by that time (which had taken some effort). I then hit "Go" a million times and kept the ones I found to come closest to what I was looking for.

The prompts I found were:

Line art, a stubby tree in a sand desert, psionic, line art
Line art, a stubby tree in a sand desert, psionic, mental energy, dream, line art
Line art, a stubby tree in a verdant environment, psionic, mental energy, dream, line art

Line art in black and white only, a stubby tree, psionic, mental energy, dream, line art, by Dave Trampier

Line art in black and white only, a tree, psionic, mental energy, line art, by Dave Trampier, by Erol Otus

Other results included:

Obviously, it simply ignored such notions as "black and white only" and it had no clue as to who DAT and Erol Otus were.

Among the main things I found, as with the marshalling thistles, was that the AIs I used had a lot of trouble linking "mundane" things with "metaphysical concepts".

Or maybe it was my prompts.

 

[overgeeked]

Obviously, it simply ignored such notions as "black and white only" and it had no clue as to who DAT and Erol Otus were.

One trick is to use an LLM to describe the art style of someone and write a prompt based on that. It only works if the LLM knows who the artist is, obviously. “[Name] style art” also works most times.

 

[ilgatto]

One trick is to use an LLM to describe the art style of someone and write a prompt based on that. It only works if the LLM knows who the artist is, obviously. “[Name] style art” also works most times.

All true. I was learning to get to grips with prompts and AIs in general and I used tips offered by various blogs and websites on-and-off. Found that much depended on which AI is used for what. Some had "categories" that sort of committed the AI to a certain style, subject, or both. So sometimes I could get the look I wanted by running a prompt through various categories and then seeing what was what.
Weirdly, that could even lead to the things generating brilliant brambles in one category but that I then had to use another category to get them with fruits on.

Ultimately, like you say, I found that AIs are not omniscient and that they have been trained for specific purposes - which severely limits their use at least for many of my purposes.

For example, I have tried literally many times to get Neural.Love and SeaArt to start drawing things in the Franco-Belgian comic style. As part of the process, I decided that generating images of city streets would be the way to go at some point and that led me to trying to make them draw a pale blue Renault 4. In all of two weeks of almost continuous trying (and then some), they didn't even come close and I was stupefied how they just wouldn't learn how to.

So that's when I found out that machines learning to do things doesn't always comply with how I think they are learning things, if at all - and what they are learning, exactly. Which is actually quite an interesting subject, for how can we get machines to understand the gist of what we're saying to them in different and changing circumstances? No tones of voice, no facial expressions, no eye movements, no moving of hands, no environmental conditions - the only things they can work with are the words you feed them and then their databases to calculate probabilities. Made me think a lot about language and how we use it.

But I digress and things may have improved since I last tried working with AIs - and will probably work better if you have your own, local AI?

 

[overgeeked]

For example, I have tried literally many times to get Neural.Love and SeaArt to start drawing things in the Franco-Belgian comic style. As part of the process, I decided that generating images of city streets would be the way to go at some point and that led me to trying to make them draw a pale blue Renault 4. In all of two weeks of almost continuous trying (and then some), they didn't even come close and I was stupefied how they just wouldn't learn how to.

So something like these? This was done with Dall-E with two prompts...

"A Tintin, Franco-Belgian comic art style. A pale blue Renault 4 on a cobblestone street."

Which barely generated one useful image. And...

"A Tintin, Franco-Belgian comic art style. A pale blue Renault 4 on a cobblestone street. One person is sitting in the car. One person is walking down the street."

Which generated these four.

 

[overgeeked]

I asked ChatGPT to describe the Franco-Belgium comic art style, like Tintin. I took the highlights from that and dropped it into the art prompt in Dall-E.

The prompt: "A Tintin, Franco-Belgian comic art style. Clean lines. Minimalist background detail. Realistic proportions. Exaggerated facial expressions and body language. Limited use of shadows. A pale blue Renault 4 on a cobblestone street."

The results:

 

[overgeeked]

"A Tintin, Franco-Belgian comic art style. Clean lines. Minimalist background detail. Realistic proportions. Exaggerated facial expressions and body language. Limited use of shadows. A pale blue Renault 4 on a cobblestone street. One person is sitting in the car. One person is walking down the street."