So many of you have seen the talk about AI generated art. There's been a lot of discussion, even on major new networks, about what AI art means for the future, the ethics involved, and the potential legal issues down the road.
I'm not going to be talking about any of that.
Instead, I'm going to give those of you who are curious a heads-up on the state of NSFW porn generation, specifically actual sex and not just nudes. For background, I'm using Stable Diffusion Automatic1111 webui version, which I believe is built on the older SD1.5 standard. I've been playing around with it for about a full day now, setting it up (which was annoying and not for the technically challenged), reading up on creating good prompts, experimenting with different models (which are basically the "pool" the AI draws its information and image data from), and creating images while experimenting with the prompts. I tried the Hassan, URPM, Smirking Face, Grapefruit, and Waifu models. I've been using extensive negative prompts, and relatively simple positive ones. So how was it?
Its not that good. Don't delete your porn collection just yet.
Its pretty obvious that the AI can't generate stuff it has little exposure to. When I tried a BJ picture under a desk, it made a few images with a desk near them, when it bothered at all. It is very good for the most part at swapping in small details, like a choker, on people, but it struggles greatly with any kind of fine details. It also has a limb problem, especially a man's third limb, as it tends to do things like making it grow double in size, float in mid-air away from the body, and duplicate it in different spots. It also struggles with actions, often ignoring explicit prompts like "kneeling". These problems applied to both realistic and hentai style art, however the hentai style was far worse. It seems that hentai models are even worse about correctly displaying dicks then RL ones, often showing a heavily censored dick that doesn't even look like a dick floating in front of the girls while they poorly do, well, something with it.
The biggest issue though, is it just flat out ignores parts of the prompt. I typed "girl getting fucked doggy style", and not one of the images showed a doggy scene. Prompt creation is a mess because you sometimes need to know the exact words to trigger creation.
That said, it does do a good job at creating good looking girls standing around or posing. Swapping in and out details like hair color and simple accessories is easy and mostly goes fine. Its great at backgrounds, and portraits. Its basically decent at anything that involves not much happening, provided you don't expect it to read your mind.
Give it a year, and it might actually be a serious threat to traditional image porn. Right now though? Not really.
from someone else. Notice it includes 'huge breasts'. But in an earlier variation of the negative prompt I also included 'big breasts' and tended to get nothing but lolis. Even though nothing in the prompt indicated loli. I guess once again it depends on the model, the images used to train it and how those images were tagged.
