CSP AI-generated players?

I have been playing with AI-generated images for a few weeks now. I was curious to see how easy it might be to create assets for CSP that way intead of using DAZ Renders. DAZ Renders can take an hour to set up a shot, and then another hour to actually render it. Not to mention all those assets cost a lot.   Turns out it certainly isn't as easy as they make it look, but gen-AI can do some pretty fancy things. I've been using Fooocus, which is free and easy to set up, and pretty quick if you have a decent graphics card.

The requirements for making a CSP opponent is that we need a bunch of pictures, in various states of undress, of the SAME PERSON, with the SAME FACE, and the SAME BODY, and the SAME CLOTHES, in the SAME PLACE. These are all things AI generation doesn't do really well.  Getting one gorgeous shot of a lady on a beach is easy.  Getting 50 more of the same person on the same beach with the same clothes, not so much.

FaceSwap and Inpainting goes a long way though.  Start with a picture of a naked lady, then use faceswap to get the same lady again in the next shot. Use inpainting on the areas where the clothes should be with extremely detailed descriptions of the clothing, generate 50 pictures, and keep the 3 that actually work out.  It almost works.  Here are a few "Vivian V3s" that worked out well, with outfits and poses matching her original beach set:

Are the shirts and the bow exactly the same? no, but close enough.  The location is somewhat different too, but whatever, beach! (also I didn't try very hard)  Is it exactly the same woman? No, but only if you look real close.  Also I had to in-paint the breasts like 50 times to get nipples that looked similar (but it was worth it).  Using PyraCanny with an original Vivian shot kept the body consistent.  I turned her into a downloadable CSP opponent for your entertainment.  Enjoy the uncanny valley.

Getting more shots of that same girl proved tricky though.  I tried a bunch but eventually gave up.  They all looked like sisters, but not the same person.

So here we are:

SAME BODY: solved(ish), but only if you have 3d rendered shots to start from, and a lot of patience with nipples.  And no tattoos.

SAME CLOTHES: solved(ish)

SAME LOCATION: solved(ish), at least for outdoor locations

SAME FACE: not really.

COMPOSITE SHOTS:  forget about it.

Pro tips:

•  start with the naked shot and add clothes, or your bodies will be all over the place
• Do all the cleanups you need on the first shot before adding any clothes, because other things change too and make it impossible to re-apply the fix to the other shots in photoshop.
• Hands suck. Be ready to redo them over and over. Or hide them.
• Don't even think of holding poker cards.  Hands are hard.
• Composite shots?  Yeah, no idea how to ever do those.
• AI gets glasses right, but the eyes behind them, almost never. Sorry Denise.

FaceSwap from CSP images did also get me a few extremely striking photoreal portraits of some pretty recognizable ladies.  The catch here is I couldn't get very many, and couldn't FaceSwap them in consistently into other shots, but still, wow:

So, after doing a whole bunch of reading and researching, it turns out you can train a thing called a character LoRA to get you a consistent character across AI-generated shots, which would solve both the SAME FACE and SAME BODY problems. (and perfectly matching nipples on the first try, yay!).  The only catch? you need to start with 30+ varied pictures of your character, with very detailed caption text files, and then run a training job that takes... 126 hours on my machine (!!!).  I have only 8gigs VRAM on my graphics card, but I'm told it would be a whole lot faster if I had 16.

Anyway, the good news is I have a whole lot of pictures of Denise, so I tried that but only ran it for a couple days with too few pictures, captioned wrong, and too many many similar backgrounds, and a mix of 3d render Denises and AI generated photoreal Denise headshots, then stopped and resumed it wrong to boot.  You don't get a lot of chances for trial and error with this.  I think I may have overtrained it too in some ways.

... It worked though!  Kinda.  Here are a few cherrypicked and heavily fixed up shots that it made for me.  (clothes drawn-in afterward.  Much, much re-in-drawing on the faces until they weren't somewhat mangled, but bodies pretty much right on the first try. Except the hands. Nightmare mutant hands, every time.

The good news:  It's definitely Denise!  And the body is spot on too, and taking poses that were not in the training data!  It's Denise's nipples too! (you'll have to take my word for it, sorry)  The bad news: aggressively having to fix the face did make her look a tad less like herself, the huge majority of what was coming out was difformed garbage beyond saving, and I had too many pictures in the same 3 rooms, so it also learnt to draw (semi mangled) red leather couches everywhere no matter how many times I asked for a beach shot.

It also never quite learnt to comingle the AI-photoreal and DAZ-render Denises, so I could get full body Daz-Denises, or closeup portraits of AI photoreal Denise. With the occasional uncanny photoreal not-quite-either-Denise in the mix. (her body before he underwear were added was phenomenal and yet still definitely Denise's:

Another almost-win here. This one started out with the dress, but I had to fight with the face quite a bit. Still, it's an almost great photoreal shot of Denise.  If I can round up a couple dozen more of these, I might have enough to spend 6 days training a new LoRA the right way!

So, it's probably time to take a step back from this crazy adventure, but I learnt a lot.  I'll leave this long post with a couple questions for you all:

• How do you feel about the photoreal AI models? Do you want to see them in the game, or do you prefer the DAZ renders?
• Does anyone have a 16g vRAM graphics card I can borrow?