What are you doing to get them to look so realistic? Especially the first angel one that you just posted. I have been trying different settings and have gotten a few that look pretty good, but nothing as realist and detailed as those.

I think the big difference is the models used rather than any special prompting. Most of the angels were made with a 50/50 blend of Wavyfusion and Elysium v1 (not to be confused with the Disco Elysium model) Most of the devils and the last angel are made with a different diffusion model blend to the other angels, with the above mixed further 33/67 Analog, just experimenting to try and get roughly 1/3 of each. They are all available on huggingface.co iirc. The Automatic1111 webui lets you blend models pretty easily, though if you're using a cloud version linked to a free google drive account it will eat up the data allowance pretty damn fast.


Any model that has been trained on custom content seems to be better at female anatomy than vanilla Stable Diffusion. Even blending in anime models with the SD 1.5 can give a strong improvement to anatomy at lower resolution outputs. For increased realism you can change to models that have more training with photography, then use img2img for upscaling.

A model trained on pregnancy content would be pretty awesome, I still don't know where to start with that and I don't think I could run it on my own hardware. But it looks like a couple of people are working on that and LadyBabyloin has posted a download link for their teen pregnancy model, in the teen ai section. I'm not sure if it's just wishful thinking but it seems to me that blending in LadyBabyloin's model helps with pregnant figures of all ages even if not using the keywords. I should do some actual experimentation to see if it really makes a difference.

Usually the prompts are unsharably cringe, but these aren't too bad.
Angels: young woman wearing angel costume, wading across a river, [pregnant] belly, cute curvy perfect body, intricately detailed masterpiece, side profile view, photo, film still,
Negative prompt: ugly, deformed, monster, hideous, bad, poor quality, blurry, disabled, chinese, japanese, asian, drawn, illustration, painting

Devils: young woman wearing an stunning demonic devil costume, waiting in the corridors of hell's dungeon, [pregnant] belly, cute curvy perfect body, intricately detailed masterpiece, side profile view, photo, film still,
Negative prompt: ugly, deformed, monster, hideous, bad, poor quality, blurry, disabled, chinese, japanese, asian, drawn, illustration, painting, huge breast

I don't dislike Asians, all the world's women are beautiful. But a lot of these AI models are heavily influenced by anime and get disproportionate representation, and I do have my preferences. I'll also use geographic indicators to guide ethnicity rather than use actual ethnicity in the prompt, it gets better results imo.

Other settings: I don't usually use the hires fix option but it worked well with these models.

Steps: 24, Sampler: DPM++ 2S a, CFG scale: 10.0, Face restoration: GFPGAN, Size: 896x1280, Model hash: a0539aaa, Model: blend_wavylesium, Denoising strength: 0.6, First pass size: 448x640

I'll try different samplers depending what feels right and the results I'm getting, but the CFG will usually be between 8 and 12 and usually at the lower end for my outputs. I'll also generate a few with different CFG values using the x/y grid extension just to see what things are like, especially if I find an output that looks like it has potential.

Finally, I'll clean up photos in an editor, or generate variations of the pics and pick out the best details, e.g. one pic might have nice tits but lousy hands, and the other might have weird boobs but human-looking hands, so I'll get the best parts together using a clone tool.

Retro Diner Date

   
   
   
   

   
   
   

I think this one really stood out from the Angels and Devils, so I've redone, retouched, and upscaled it.

   

You're amazing at making these!

I know. But the AI does all the work, and the AI does what it's learned from the artists, photographers and models whose work it was trained on, so it's a team effort really .

 AI can be directed to make specific scenes by experienced artists but I can't do that yet. For now I don't make these pics so much as discover them.

   

   

Merry Christmas Everybody! Gonna call it quits for the rest of the year.

   
   
   
   
   

Happy New Year!