New implementation of my last remaining old mix. Here the core is composed of Kimura Takahiro style and a selection of furry art. Quite important if used in combination with toon, disney or similar styles: if you use high weight and you do not want furry traits, specify "furry, not human" in the negatives: Disney often used anthropomorphic characters, and the furry style can resurface in those cases.This is quite more sensible than other models to tags to flatten the colors, like 2D etc. With the impostations about gamma and noise, this resulted in a quite distinctive color style and some versatility between not-realistic and realistic.The Adam version is the less intrusive one, I think: now I get all the times I read Prodigy produces rigid LoRAs, that one tends to modify the prompt more.

Flat Mashup

Prodigy version. More or less to compare the influence of the optimizer on the LoRA. Even with minor size!

Flat_mashup_-_second_coming.safetensors

E10 prodigy

ADAM E11

This is more or less to confront a LoRA made with AdamW8bit with a LoRA made with Prodigy. Same dataset of the other Flat Mashup, so a lot of Kimura Takahiro, a good quantity of furry art, something else. Probably this tensor will end as another version of that LoRA.

Files with SHA256: 63fcce205859c6d1f2c15da29b5a110c97e5c1bbe1e2b586d7a701620d039d9c