Some news.

I've been quiet lately, but that doesn't mean I wasn't working on new stuff. Actually, it's the opposite - I've tried so much new stuff that I didn't really have time to share my findings.

There won't be any new SDXL versions. It's just not worth it to spend more time, trying to squeeze out more out of a really outdated architecture. I was working on Chroma finetune, but then Z-Image came out and made it pointless to work on Chroma as well - Z-Image is noticeably faster, has higher quality and better aesthetic and an actual LLM as a text encoder (technically speaking T5 is also an LLM, but you get what I mean).

So, right now I'm waiting for Z-Image base model release, because there's not much you can do to Turbo version in terms of finetuning. Refining my datasets, gathering as much 2k resolution images as I can (finally we can make use of them), redoing captions. Also, improving the variety of concepts in the dataset - ZIM's 6b parameters would allow it to adapt nicely.