stable-diffusion-webui-feature-showcase VS stable-diffusion-embeddings

A1111 https://github.com/AUTOMATIC1111/stable-diffusion-webui-feature-showcase

Here's a quick rundown on a few of the most popular ones with links. I started out using CMDR2, which is very easy to get running as a newbie. Then I kind of graduated to NMKD because I wanted something a little more mainstream but still easy to use. Then, I finally decided I was hungry for all the strange and exotic bells and whistles that SD had to offer me, and so I installed Automatic1111. I also wanted something that would work well with my 4GB GTX 1650 laptop card, because that's considered "low ram" and kind of on the edge for running SD- Automatic1111 fit the bill there, too.

All generations were on the "Berry's Mix" model, which is made by combining NAI-final, Zenith's F111, r34 and SD1.4 according to this recipe. I used 30ish steps when generating images and inpainting, but 70-80 steps when outpainting because I read here that outpainting really benefits from extra steps. When outpainting I would generate 2-4 versions and pick the least broken one, then tidy up with inpainting.

sounds like "outpainting", one of the very first features listed on 1111 repo with some instructions: https://github.com/AUTOMATIC1111/stable-diffusion-webui-feature-showcase

I have no idea what you're talking about. Just get Automatic1111

My own dataset consists of 107 high resolution input images and I attempted training it for more than 50k steps mulitple times with varying parameters. One would assume that it should work; AUTOMATIC1111 trained his embeddings here quite similarly, but my own results looked quite plastic-like.

Warning: NSFW https://gitlab.com/16777216c/stable-diffusion-embeddings

Yeah, that's what i've been training on but the step count was 3k to 7k not 100k, with that many steps you might need more data but i'm not sure. Check Automatic's TI wiki, those are the ranges he's using, there are also some nsfw embbedings on gitlab trained like that.