lol.. you know when you've been using dynamic prompts to much when..

@epinikion It is easy to overtrain because it seems to be WAY more cleaner. They culled lot of SEO spam/clickbait from the training process which helps a lot. Adding lots of precision for everything. Downside it that because there is literally less shit way to dampen the process of training/finetuning/making LORAs whatever, it is way easier to overshoot. My limited experience with SDXL really seems to indicate that whatever you can use to dampen the training process you should use. Because it amplifies things very quickly.

1.x and 2.x had the problem that many common tags were way too polluted by SEO spam/clickbait that certain tokens were just useless. Now captioning is very important and needs to be accurate because every term you use represent more what you actually think it means. Every new iteration of the base models AND derivative models in my experience has added multiplied the quality every time. There are models on civit nearing their 10th iteration just on SDXL, but I have noticed some of them have started to become more rigid as in able to only make that thing the derivative model is meant for. Which I guess is good for most people who just want that thing.

You can inpaint on Fooocus with any model flawlessly.

Or if you use ComfyUI you can use the https://github.com/Acly/comfyui-inpaint-nodes nodes to have the same effect as Fooocus