D&D General DALL·E 3 does amazing D&D art

[Parmandur]

The training set must be absolutely stuffed full of beautiful female elves. That’s testing the generator on its strong suit.

Try to get it to do an ugly female elf. Or just a plain, average, slightly plump middle aged female elf.

Well, yeah, exactly. The algorithm is good at following tropes that are well trodden. You can get thia wild stuff by following that trail, bit of you push against thw wnvelopw you find Weird Things (like the insistence on giving Hobbits big bushy beards, or the inability to do female Dwarves). There is a lot of garbage the system throws up.

 

[SakanaSensei]

The training set must be absolutely stuffed full of beautiful female elves. That’s testing the generator on its strong suit.

Try to get it to do an ugly female elf. Or just a plain, average, slightly plump middle aged female elf.

Pretty much any female character that isn’t conventionally attractive feels like pulling teeth. Which says a lot about what kinds of art is available for it to pull from in the first place unfortunately.

 

[EzekielRaiden]

The training set must be absolutely stuffed full of beautiful female elves. That’s testing the generator on its strong suit.

Try to get it to do an ugly female elf. Or just a plain, average, slightly plump middle aged female elf.

It's also deeply in love with tieflings. I've asked, numerous times over, for "dragonborn paladin" and other such things. At least half of the images I get are actually tieflings with draconic wings.

 

[Reynard]

Good morning.

Gives me Middle Earth vibes.
Prompt: glorious sunrise in green hills where an ancient giant strange statue lies broken amidst small hovels, digital fantasy art

 

[EzekielRaiden]

Good morning.
View attachment 297359
Gives me Middle Earth vibes.
Prompt: glorious sunrise in green hills where an ancient giant strange statue lies broken amidst small hovels, digital fantasy art

Pretty sure it's pulling from the Wheel of Time here. That's the male Choedan Kal, just without the orb in its hand.

 

[Oofta]

Most of these images are truly amazing, but I still can't help but notice the goofs. I understand why it happens but take this pic taken from this this with images that @Parmandur generated.

The dogs and riders in front are amazing, the castle has too many minarets but other than that not bad at first glance. Then I look at some of the detail

Those gnomes have been doing some serious cross-breeding there to give a horse dog legs.

WTF is going on here?

Generated pictures that have 1 or 2 creatures seem to have fewer obvious glitches, but when you start getting these crowds the background images almost always have serious glitches. Doesn't make the technology any less equally amazing and terrifying of course, just that there is a limit to what it can do because it doesn't really "understand" what it's doing. It's just recombining art from scraped sources (which has it's own ethical concerns), throwing it into a blender and reassembling based on recognized patterns.

 

[Reynard]

Most of these images are truly amazing, but I still can't help but notice the goofs. I understand why it happens but take this pic taken from this this with images that @Parmandur generated.

View attachment 297361

The dogs and riders in front are amazing, the castle has too many minarets but other than that not bad at first glance. Then I look at some of the detail

View attachment 297362
Those gnomes have been doing some serious cross-breeding there to give a horse dog legs.

View attachment 297363
WTF is going on here?

Generated pictures that have 1 or 2 creatures seem to have fewer obvious glitches, but when you start getting these crowds the background images almost always have serious glitches. Doesn't make the technology any less equally amazing and terrifying of course, just that there is a limit to what it can do because it doesn't really "understand" what it's doing. It's just recombining art from scraped sources (which has it's own ethical concerns), throwing it into a blender and reassembling based on recognized patterns.

If I understand the technology correctly, what generative algorithms do is "decide" what the next piece is based upon the training data for the prompts given. It isn't a collage, but instead a blind person painting something based on a massive spreadsheet of cross referenced pixel codes. That it works at all is amazing.

These artifacts are good, in that they make it easy to spot AI art and that means people can make informed choices about whether they support a thing that uses it.

I think that art to train AI should be opt in if it isn't in the public domain, with whatever agreed up relationship the AI company and artist cone up with. It isn't going away and has legitimate commercial and cultural value, so we might as well use it for good rather than evil.

In the meantime: more gnome doghorses!

 

[Tales and Chronicles]

I managed to generate pretty good large-sized women. I went for middle-aged, but I'm not quite sure its there yet, unless middle-aged is around mid-30s.

 

[Reynard]

I managed to generate pretty good large-sized women. I went for middle-aged, but I'm not quite sure its there yet, unless middle-aged is around mid-30s.

View attachment 297369View attachment 297370

What did you use for a prompt here?

 

[Tales and Chronicles]

What did you use for a prompt here?

I tried to insist on the different body shapes, making sure it ''understood'' what I was going for:

''A middle-aged plump, curvy, large woman plays the lute in a town square. Sketch line art.''

I'm working on a PHB redux for my table(s) and I have fun re-creating the class imagery. Here I went with a a beautiful large women bard in a calm setting to oppose the tropey slim bard womanizer in a tavern setting.