Prompt: ugly humanoid with pot belly, tiny closed eyes, has pale gray skin, shark pointed teeth, pointed ears, long patches of dirty hair, wearing a loincloth of animal hide, armed with a spiked bone club, primitive caveman, powerful build, wild animal, low angle camera shot, high detail, digital painting
So, my wife asked me if I could make an image in the style of Maxfield Parish but with a wider range of body types. Growing up she never saw photos or drawings that looked like her. I said I would try because we all know that DALL-E can get weird when you try and deviate from skinny, athletic women.
There was a whole journey that got me to the first image (I cannot tell you how many nude and topless women it created, unasked, for me).
I am dialed in right, so I ask for an edit where I change the ancestry from human to teifling (with green skin and horns) and instead of keeping my subject well rounded it does this...
Now, I am simply telling it to use the previous prompt from the image but change one thing. Evidently that means default to skinny and athletic (or someone with an eating disorder). It is frustrating because I have to be ultra careful how I ask for someone with a bigger body size or I get the dreaded...
So, I gird my loins, and I try to get back to where I was before... First thing I do is copy a working prompt with a human woman and then hand edit it instead of natural language prompting for changes. I input the following:
A digital painting, close-up portrait of a woman of tiefling ancestry with a fuller, heavier body type in a diaphanous gown to showcase natural and asymmetrical body features, portrayed in a vibrant landscape. Her skin is light green and she has dark green horns highlighted with black accents. She showcases natural and asymmetrical body features, set against a vibrant landscape with brightly colored skies. The teifling, in a relaxed pose, is gazing into the distance, surrounded by lush, detailed vegetation and whimsical architectural structures reminiscent of classical antiquity. The artwork combines luminous, rich blues and vivid hues, focusing on a luminous effect with glazing techniques, evoking romanticism and idealism. Fantasy artisan prompt, hd quality, natural look --ar 1:1
Not quite what I am looking for. Using the last image as the basis for the next instruction I say, "Still too skinny." That generates this:
I tell it, "Closer, make her face and body rounder, less toned."
Not sure if this is neutral or step back so I say, "Rounder."
Ok, now we are getting closer. A little too cartoony for what I am going for so I have it regenerate twice...
Now we are cooking with fire. Using the last image for reference, I tell it, "She is a sorcerer with a bird of paradise familiar." Which gets me to this image... that is a lot of work to get what I want...
Postscript: I have left out a whole journey to figure out what words I could use to direct DALL-E to the image I wanted. That was a days worth of work in a completely different conversation than this one. From those lesson learned I took the best of three Generated Prompts to create the one that starts this conversation off (which had the added benefit of being safe for work). There were minor tweaks (changing the Aspect Ratio for example) and but for the most part is was using Natural Prompts to get it where I wanted it.
Prompt: ugly humanoid with pot belly, tiny closed eyes, has pale gray skin, shark pointed teeth, pointed ears, long patches of dirty hair, wearing a loincloth of animal hide, armed with a spiked bone club, primitive caveman, powerful build, wild animal, low angle camera shot, high detail, digital painting
Just a FYI, you can get cleaner "Thumbnails" of your images by going to each attachment and adding back in the thumbnail version of it. Great work on the monsters by the way.
It's not as immediately obvious as the Empire waistline thing, but the AI is also really, really bad at saris. It knows there's a choli (a blouse) and that there is flowing fabric that drapes, but the saris it produces don't actually drape in a way that makes sense, and it's often hard to figure out exactly how it connects. Sometimes it just looks like big sleeves.
Which makes sense - how does AI know what clothing is in the first place. I should stop trying to reproduce specific fashion, but I live in hope.
Dungeons & Dragons book cover high quality oil painting, a female 6'2" tall 220-pound statuesque strong athletic robust long-black-haired grey-eyed pale-skinned Celtic woman in well-fit leather suit with a lacey white high-collar button-up shirt with lots of jewlery and a sword on her back walking through a fantadtic steampunk city with a spider theme
It's not as immediately obvious as the Empire waistline thing, but the AI is also really, really bad at saris. It knows there's a choli (a blouse) and that there is flowing fabric that drapes, but the saris it produces don't actually drape in a way that makes sense, and it's often hard to figure out exactly how it connects. Sometimes it just looks like big sleeves.
Which makes sense - how does AI know what clothing is in the first place. I should stop trying to reproduce specific fashion, but I live in hope.
