Improving Accuracy of Image Generation
Added 2025-02-19 00:11:30 +0000 UTCI've been focused on improving my image generation. Sometimes I can get great things through Stable Diffusion, sometimes it seems I can't get anything but horrific messes.
First off, I realize I was using weights wrong. Square brackets [] is for negative weight reducing the prevalence of the thing. I was using it similar to parenthesis () which are used as positive weights. Also, I was using too heavy of weights. PonyRealism doesn't seem to like weights above 1.5, and the images I was producing were getting messy and deformed until I stopped.
Secondly, I'm looking into doing InPainting/OutPainting and Regional Prompter to try to improve things. InPainting allows you to specify an area an image to "touch up". This is key for when I have a great image but the face or hands or genitals are messed up. OutPainting allows you to expand an image. And Regional Prompter should allow me to better handle multiple characters/subjects and compose scenes. Right now, scenes I'm working on with multiple subjects are a mess, with Stable Diffusion not able to figure out which part of the prompts belong to which person and constantly swapping them out. Sure, you can push through with lots of generations, but it can be hard to get good images. See some of the example images I've been working on for a sex scene. And these are some of the best after generating over 300 images. It is supposed to be a thicc muscled futa with black straight hair in a bun fucking a chubby, fat MILF, with brown curly hair, but traits keep getting swapped around or ignored.
