In Midjourney I did try to add things like --no deformed, disfigured, malformed or hands have exactly 5 fingers each in their correct anatomical places to the prompts but with no success.
Do you have any tricks to make those models generate plausible looking hands?
I’ve learned to shrug, roll my eyes, and laugh at the 17 fingered hand.
Gotta love that horrific blobby mass with fingers jutting out and twisting at all angles.
I agree, I’ve stopped counting fingers a long time ago. It’s just part of the AI esthetic.
I’m happy as long as there are less than ten fingers on one hand.I’m happy as long as there are less than ten fingers on one hand.
Man, and I’m just happy as long as I get one hand per hand, regardless of the fingers!
Control Net on Stable Diffusion has a “hand and body” pose. You can individually place fingers if you use some other extensions, or you can find a real-life picture of someone whose hands and body are posed the way you want them and use that as a template. That might be over-kill, though.
At this point though, you might as well skip AI and commission an artist (photographer, painter, fx).
Negatives help, you can also specify hands be concealed behind objects or whatnot. There really is no reliable method except training larger models, (eg SDXL gets it right a lot more often than standard SD models) or manually editing the image afterwards.
First I try to add +25 or +50 steps in the generation. Sometimes the limbs, hands, face are self corrected.
When the size/résolution of the hands or of the face are too small, I can’t get what I want. The results are shitty whatever I try.
Also, the models aren’t equal. Sometimes, I inpaint the things I want to correct and launch it with another model. It still does work well if I try on very small faces.
