I posted about my AI porn site pornpen.ai here last year and it reached the top of the front page. And yes, it's still going strong :D (and we've integrated SDXL and videos recently)
I think what happens is that in the early stages of generation, the model picks up on arm-like structures and just keeps building it out. It doesn't realize that the overall structure will be flawed later on.
Creator of pornpen.ai here, looks like this discussion is ramping up again! Wanted to respond to some of the points made here and in the article.
The grotesque images: Simply put, the model is not good enough yet. This will improve over time as the models get better.
Diversity/male gaze/heteronormative: Generating men and other genders did not yield very high quality results unfortunately, so I didn't include it in the initial launch. I am also manually fine tuning the prompts, so it was easier to pick one category (women) and go with it. I hope to expand this in the future.
Sex workers: I do agree with the quote in the article that this will not replace sex workers (at least for now) in the same way that I don't think dalle/midjourney/stable diffusion will replace illustrators. In the future, I believe this technology could create a new category of sex worker that can generate content without using their own body (similar to VTubers).
It's harder to say what will happen in the future though. If these models keep improving at this pace, I think these types of discussions will be more important. Let me know if you have any questions!
What are your thoughts on also releasing the prompts and maybe even seed for each image? I assume it wouldn't be much work and would allow people to fine tune them.
Good question. I figured that many people who come to the site will just press generate right away without realizing you can choose tags. When 0 tags are selected, you get weird results. So I wanted to make a good default selection. But maybe a better option is to force users to select some tags first, and start empty?
FWIW, I think your initial idea is the better one. Like you said, people will just want to hit Generate to see what happens. Modals in the way will just annoy. Maybe just randomizing the start tags, for variety in the stream?
I think the term “perfect” is a bit abstract/subjective as a qualitative assessment. Perhaps make it a bit more descriptive in an objective way what kind of type of body it is?
Maybe start with just male genitalia. I hear some males have the tendency to send this unsolicited to females. It would be easy to build a data set around this. Just put out an open call on social media. Be sure to specify 18+ and log all users submitting.
hahaha. I can see a future of unsolicited non-existent dick pics being sent to people.
for that matter, the catfishing aspect for thirsty people make this an interesting evolution. What a strange and fascinating world we're building for ourselves here.
"The image classifier said this image is 97% likely to be a dick pick, and has automatically muted it. Click here to show anyway. Click /here/ to keep it hidden but reply with a machine-generated bigger one."
That's a way to classify whether or not the pics are indeed dick pics, but the problem lies in how these recent diffusion models work. Typically you have two encoders, one for the text prompt and one for the image. Then you train a model to optimize where the prompt embeddings match the image embeddings. You've got to actually have some decent descriptions of the dick, like: thin, girthy, 7", 8", 4", micropenis, veiny, limp, hard, huge tip, bifurcated, foreskin, black, white, latino, asian, etc, if you want to generate with any accuracy your desired fake penis.
I posted about my AI porn site pornpen.ai here last year and it reached the top of the front page. And yes, it's still going strong :D (and we've integrated SDXL and videos recently)