Ugh... I think this one is overbaked. I clicked publish so I could test it without realizing this meant I lose ALL my prior epochs, and I've rarely actually published the tenth epoch from my usual runs, as it get stuck on minutiae and loses the "big picture" imo.

This is effectively a test of a method I use in SDXL to improve prompt adherence. The flipside is a form of ablation, but I haven't trained that "partner" model yet, as it's a challenge to test on Civitai, and my local computer cannot run the loras with the quantized flux model.