This is mostly still in-need of real alignment. Right now it's less of a working graft and more of a aesthetic structured noise injection with donor model as prior. Doesn't mean the outputs arent kind of nice, but eventually I'll revisit with probably Zimage and much more depth bias to avoid compounding error after first layers which is graftings biggest issue with bad alignment.
This is an attn Q + K and MLP gate up groom using Krea2 on MOSS sft. It does not add data or weights and it doesn't make the model output like Krea2. It changes behavior in attention influenced by Krea2's attention like the other grafts. Causing MOSS to tokenize ever so slighlty like Krea 2, resulting in mostly more details, and usually some large improvements in background compositions. It's like having a subtle detailer-type Lora and levels adjustment Lora across the model which doesn't interfere with any prompts or seeds, but nothing was trained and training wasn't needed. Excitement on the v band was lessened and merged lower but it still introduces more contrast and saturation which I left in because I did feel like sometimes the base MOSS model was overly desaturated and flat by default. Composition and character details are the most significant departure, technically none of what is output has anything to do with Krea2 although I have noticed some Krea2 behaviors like how it does earrings occuring, but this is mostly MOSS's attention being almost distilled/groomed on itself to mimic how a different model attends tokens. Key advantage as always being entirely non-destructive in as far as the grafts go, training-less, "impossible" if you ask "educated" elitist ML nerds, but end result is a pretty much a general improvement over the int8 sft on a lot of unrefined prompts that don't work well on base, a lot more sharper and more colorful art/anime gens, pretty much universal improvement in background compositions, and good parity with the first versions of the DMD turbo allowing it to be used at less strength like 0.7-0.8 with 8-12 steps.