I generated over a thousand supermodel images with SD. Here's what I found

I just spent the past 48 hours to generate over a thousand model images with Stable Diffusion.

I took the best 5% and released them on https://www.thesemodelsdonotexist.com/.

You can use these images for any project.

Takeaways:

- Stable Diffusion (SD) doesn't understand/can't generate hands (huge issue when it comes to human models) - Negative prompts many times didn't work for me (if you have figured this out please let me know) - Certain races are easier to create than others, the model certainly has bias. White male models are easy to generate whereas I really struggled to generate Latin-American women. I didn't quantify the bias - this is only a feeling atm. - GFPGAN works very well with SD images if the face doesn't look good, but if it does look good GFPGAN usually messes it up. In these cases I used a simple upscaler to scale up the images. - If you change the aspect ratio (I used 512 x 768 pixels) stable diffusion can do really weird things, very often it duplicates the content. In my case, it very often placed two humans on the photos - again, if you figured it out how to avoid this I'd appreciate if you shared this.