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

[Kichwas]

I started messing with HiDream last night. So far I have been unable to make anything non-caucasian. That was a general problem in all AI Art about 2-3 years ago.

I haven't yet tried to load it up with ethnicity and skin tone loras that predate it by a few years, but that's what I'm going to try today.

My attempts has a lot more trouble making the weapon a 'maul' or great hammer, and it's so far got a strong style lock. It looks worth pursuing because it has less desire to melt fingers together or add too many of them...

I don't know what kind of weapon this is, and it probably would be just as useless as a Klingon bat'leth (for a similar reason - the way it's held removes the 'force multiplier' you'd normally get from a long shaft), but I want it on a character now.

The keywords for this one were:

"A vibrant 1970s fantasy style RPG illustration of a freckled dark brown-skinned African dwarf woman, age 29, in a fullbody shot. She wears a platemail breastplate, and her black wavy hair is tied into a ponytail, adding to her fierce and majestic appearance. She holds a giant fantasy warhammer (hammer, maul) in both hands, standing confidently in a bustling underground dwarven cavern fortress. The cobblestone streets are lined with lively market stalls, filled with various goods from weapons to food. The underground town is filled with dwarves in soviet attire, engaging in various activities, creating a lively and dynamic atmosphere. The lighting is moody, casting a deep tone over the scene. The dwarves are seen bartering and socializing, with the overall atmosphere being one of camaraderie and industriousness. The dwarf woman has intricate rune tattoos. The camera captures her full body."

And yeah, I added 'dwarves in soviet attire' as that seems like, if I can get it to work in here, would be a bit of a funny add in to the theme.

She doesn't exactly look 'dark skinned' to me though. And she's still not stocky enough to be a dwarf.

This project of mine started solely to make a token for an upcoming one-session long one shot character that won't even be seen again after that - but it's interesting and if I can get more variety out of HiDream (which I hadn't heard of before yesterday) that would be really nice.

 

[Nightfly]

These programs clearly can't handle female dwarves, and really can't handle non-white female dwarves.

 

[DragonLancer]

The problem probably is, that there isn't enough female dwarves in artwork it has sampled and it doesn't know how to apply an African ethnicity to them because of it.

 

[overgeeked]

It also takes awhile to train the thing in unusual combinations. I default to dark-skinned women when writing prompts so it's more used to pulling from those kinds of references.

The woman's description is: "A dark-skinned woman stands guard. She's wearing armor and carrying a battle axe. She has a broad-shouldered and well-muscled body." That's all it took to get the different skin tones, different hair styles, and different facial features.

Every time I told it to make a female/woman dwarf it spit out a heavily-bearded guy.

 

[Jfdlsjfd]

Well, if female dwarf isn't understood by the model, we can tell him what we want instead. I think we need it because for us a dwarf is a short and stout human with beard, so we know that any description that apply to a human can apply to a dwarf. It is possible that the AI stores the data about dwarves in a place that is closer to lightbulbs than humans, so an African dwarf could make as little sense to it than an African lightbulb to us.

Using HiDream locally, I got "more dwarfy" proportions by fumbling with prompts like that:

A short and stout woman, african face, with a ponytail and maeup and earrings, holding a huge maul, and wearing platemail.

It is also possible that the way the images are captionned explain the difficulty to produce some ethnicity. For example, I am pretty sure the captionning used is never "a brown skinned person" as you used. It is more probably labelled as "an african man" but north-african can also be labelled as such, and they are white. So it might confuse the AI which eanr that "african" means often black, but sometimes white.... They are even possibly labelled as "an african-american" or "of african descent".

I tried an african-american dwarf in the prompt above and got this:

Which is slightly a darker skin tone than african alone.

I am pretty sure loras will be available soon for HiDream. It's a new model and wasn't finetuned a lot yet.

 

[Kichwas]

I had to pile on a stack of old LorA's to get to here, along with prompt changes:

"A vibrant African fantasy style RPG illustration of a short black African woman with dark brown-skin as a short fantasy dwarf woman with a stocky build, age 29, in a fullbody shot. She wears platemail armor, orange and yellow fabrics, and her black dreadlocks are tied into a ponytail, adding to her fierce and majestic appearance. She holds a single giant fantasy hammer as she stands aggressively in a bustling underground dwarven cavern fortress. The cobblestone streets are lined with lively market stalls, filled with various goods from weapons to food. The underground town is filled with busy dwarves in fantasy attire, engaging in various activities, creating a lively and dynamic atmosphere. The lighting is moody, casting a deep tone over the scene. The dwarves are seen bartering and socializing, with the overall atmosphere being one of camaraderie and industriousness. The black skinned dwarf woman has intricate rune tattoos. The camera captures her full body. ((dark skin)) BattleHammer, Hammer, Holding Hammer, Fantasy Hammer. Her left hand is in a fist."

I've essentially spammed it with ethnicity and skin tone descriptors because just saying it once means it presumes I actually meant an Irish or German woman... (which is still better than over in the Stable Diffusion tools that kept trying to insist I meant Chinese anime when I'd type in "Mexican woman playing guitar" or something ).

But it seems to want to put her in a soft top instead of armor now... And I've been stuck with 1-handed weapons or that weird double headed polearm. I'll try 'huge maul' next.

Of note is that this is still leagues better than what I get without HiDream. Most of my AI stuff takes hours of hand drawing in art tools after it's made, and then dozens of passes again in the tools. I still would do that with permanent characters - but the hand stuff would be refinement and I wouldn't need 100+ images before being able to pick just 1 to customize.

At this stage, I'm tempted to try this with other character ideas also, as it is taking less to get to promising results.

 

[Jfdlsjfd]

I've essentially spammed it with ethnicity and skin tone descriptors because just saying it once means it presumes I actually meant an Irish or German woman...

Nice, I think the end result is quite close to what you had in mind, even if it's not close enough yet.

Of note is that this is still leagues better than what I get without HiDream.

This is one of the best models available currently. But since there have been a series of releases over the last six months of large models, that take time even on top-notch gaming PCs to train (or requiring hiring cloud computing), none of them had enough staying power to get the same level of focus as SDXL.

ChatGPT is great, but is too censored. It won't draw blood or battle scenes with more than just to characters sizing each other, it will need a lot of prodding to draw dead or destroyed things... But these are often needed given the violent nature of activities PCs are often involved in.

SDXL has a lot of varieties and RPG-themed finetunes, but it's old and superseded by larger models;

HiDream is currently the best open source one and still fitting on a gaming PC at home.

SeeDream (totally unrelated to HiDream) is an online Chinese model that is very, very capable and with 60 generations per day, it's free-ish and is currently the best quality available for RPG needs. It does OK violence.

And, sometimes, even adds blood where none is warranted:

Prompt: a Jedi knight sommersaulting over the carcass of a defeated combat robot.

Most of my AI stuff takes hours of hand drawing in art tools after it's made, and then dozens of passes again in the tools. I still would do that with permanent characters - but the hand stuff would be refinement and I wouldn't need 100+ images before being able to pick just 1 to customize.

At this stage, I'm tempted to try this with other character ideas also, as it is taking less to get to promising results.

Nice! I am with you that AI is great for quickly generating tokens and "disposable NPCs" but it still needs work for permanent characters unless you're ready to be very tolerant.

 

[Kichwas]

Nice! I am with you that AI is great for quickly generating tokens and "disposable NPCs" but it still needs work for permanent characters unless you're ready to be very tolerant.

For my own characters I'll often end up making almost 100 images, and running them back and forth between the AI and Gimp.

Last night I put a lot of my own characters through img2img with hi dream and the results were mixed. But I see potential.

I was on Automatic1111 a day ago, and am now learning ComfyUI which gives me access to both Hi Dream and Flux. And I think in that range are tools that would be most useful to people wanting 'quicky character art' for RPGs. Flux has some tools for some non-humans, and Hi Dream is likely to get there soon.

But the install process and system reqs are beyond most folks. My players tend to hand me tokens made with Microsoft Designer because if you're not too picky it usually has a usable result within the first few tries. Great for one-shot games where you just want something other than a 'default token' image on the VTT screen or as a little printout on the table in front of you.

For a full campaign, especially something like a liveplay - hire a commission.

 

[Scribe]

These programs clearly can't handle female dwarves, and really can't handle non-white female dwarves.

Took me many many attempts to a get a female dwarf, its not great at other body types.

 

[dave2008]

A lot of the AI art in this thread is superior to the crap art in the 2024 books IMO.

Do you really think the 2024 core book art is crap or do just not like the style or theme of it? I can not like an image, an artist, or style and still realize it is not crap. It is just something I don't like.

Though for me, whether I like it or not is completely besides the point when it comes to AI art.