This post will be an experiment with the main character on my current run done with different AI engines, models and styles. Zeno_42 has done similar posts at his excellent Zeno’s Ziggurat site, things that may be more broadly useful and interesting! But I sort of want to see where we’re at now, in January of 2026.
This post is inspired by a comment at the Beamdog Forums, it inspired me to want to try a few different things.
These will all use the same prompt, with any minor changes or tweaks mentioned as we go. Gabby is a good natured, high energy sort, a fighter (7th level) dualed to Cleric of Lathander (sunburst is a symbol of Lathander). I should mention “realistic photo” is *my* choice of style for BG artwork. It is what I like. I understand many do not share this view, I will try some more traditional sorts of art too.
warrior woman (age 20, Slavic, fair skin, high cheekbones, short, cute, blue eyes, shoulder length pale blonde hair, wearing silver plate armor with gold sunburst motif) stands with hands on hips and smiles with raised eyebrow, Standing in a courtyard near a castle wall; realistic photo, medieval fantasy
Bing Co-Pilot DALL-E 3. My favorite renderer, I think its a nice balance of realistic and colorful/fun. Bing Co-Pilot MAI-Image-1. Beautiful crisp and realistic image, but why does it make her 12? And it does this consistently. It seems to do this less with male characters? An odd flaw.Meta’s OpenArt most current “Nano Banana 2” renderer, this is shockingly better than I expected. Drab colors, but sharp, clear, realistic. Maybe promising?Meta’s OpenArt with my own “Gabrella” model. This was built from several of my favorite Gabby renders from other renderers. OpenArt does have some other limitations, but I love it for simple portraits.Meta’s OpenArt with their “DnD” model. This is vastly better than the last time I played with it.Now I want to change up the style of the render a little. Going back to my base DALLE- 3 renderer, I swapped the “realistic photo” style cue with “Clyde Caldwell style fantasy art”. Its not hard to see why real artists hate this stuff. But I’m amazed, it literally took me longer to change the text of the description than it did to render four examples.The same as above except I changed it to “Larry Elmore”.Keith ParkinsonMike Sass, who did the original portrait art for BG1. I’m not convinced the AI has any idea who Mike Sass is, it just did “fantasy art”.Yes, its anime Gabby.Justin Sweet did the portrait art for IWD. I added the word “snowy” to the scene description.
I’ll have to play with OpenArt a bit more. I had largely given up on its ability to handle complex scenes with motion and multiple characters, but its been a while since I gave it a chance.
I suspect “cute” is what’s making MAI-Image-1 turn her into a tween. Certain words just really bias age – especially for women. “Mature” is another one that can quickly override the numerical age.
I still use “Luis Royo” an awful lot lately, but series styles like “Dragonlance” or even “D&D Art” sometimes help as well. “Justin Sweet” was what I started with when I first got into all of this, but I was never quite convinced that it actually knew what that meant. I’ll have to experiment with him again. “Wayne Reynolds” can also be useful for a bright, colorful look (he did a lot of art for Pathfinder). And “Jeff Easley” was my go-to for several runs – like Roland in Solasta.
The main issue I tend to have is finding style descriptions that hold up consistently through an entire campaign. Something that looks good in the frozen north may look quite different in a jungle setting. I often end up with something like “A Luis Royo dark fantasy painting” as my basic cue – just to nail down both artist and genre.
OpenArt has so many Models to choose from, I suspect some of them may be very good. But of course, Choosing on of their madels makes you unable to chose your own. Although now can use both a model with a reference photo. Which all makes me think OpenArt might be very powerful if I played around with it more.
I had sort of concluded the same about “cute”. Especially with “short” in there too! Its one of those things that doesn’t screw up all the time, just most of the time. But then it also makes Bull too young just not as dramatically, like 30 instead of 40.
I stopped saying D&D Art because it tends to add d20s all over the place! But I do add “Fantasy art” or “medieval fantasy”, pretty much always. I would agree it doesn’t actually know all the artists. Which confuses me a little since I know they can use the internet. Its right up there with the mystery of crossbows. I was talking with an AI pro friend just the other day who was amused to hear some of this. His answer is that AIs vary a lot depending on what all they’ve been trained on. And this should be improving, probably pretty quickly.
I’ve also found the same about style strengths and weaknesses depending on subject. I have done a couple of IWD runs with all the same artist, but most often I switch at some point depending on what I’m looking for. Even with “realistic photo” as the cue I get a shocking number that look more like paintings or drawings. Some of that makes sense, how to you show a photo of something fantastical? A dragon? How do we know its a photo or a drawing? Yet I guess our brains look at everything from detail, to color and shape consistency, to backgrounds and can make a judgement on if it looks “real” or not. AIs are not quite as nuanced yet!
DavesGaming 
Leave a comment