Wuerstchen VS stable-diffusion-aesthetic-gradients

Fully correct, also the v2 of the paper introduced a model that is bigger and slower, however generates better images. So the 500ms was only for the first model we introduced in v1. I also want to mention our new work as it is very much related to this whole topic of "speeding up models" -> either training or sampling: Würstchen: https://github.com/dome272/wuerstchen/

You might want to try https://github.com/vicgalle/stable-diffusion-aesthetic-gradients to see if you can reproduce the style without necessarily having to come up with text descriptions

Aesthetic gradients plugin for Automatic1111 WebUI

I think you're describing Aesthetic Gradients

Its different imo. Look at the floral example flower_plant https://github.com/vicgalle/stable-diffusion-aesthetic-gradients

Also, since you have found some interesting aesthetic embeddings, if you wish, you can contribute your embeddings by opening a PR in the repo https://github.com/vicgalle/stable-diffusion-aesthetic-gradients !

We could definitely use some help with generating aesthetic gradient tensors if you think you'd like to take some of that on. It's a fairly straightforward process and it's all laid out in this repo. I would love to see what you come up with and how you implement it. The file output is a pytorch tensor so it's extremely small which makes it great for sharing with friends!

You could look into generating your own aesthetic gradients using a dataset of people from a distance with faces you like in the style you're going for. Here's the repo. It's much more compute efficient than retraining over and over to adjust for overtraining with many tokens at a time. Although it can take some time adjusting your aesthetic steps. You should be able to get results! Just explore and have fun!