TL:DR
It's a Style LoRA which attempts to focus on a slightly more realistic style, but doesnt quite manage to do that, instead going for a unique style that looks, well Unique.
Does it have preferred settings?
Test it out for yourself, I've gotten decent results with most settings, but generally i'd avice going for:
Just 1 Character at a time (1girl, solo)
A median CFG value (4-20),
Sticking between 0.4 ~ 0.8 LoRA strength.
Steps don't matter too much unless you are using an ancestral Sampler (In my experience)
Model was Trained on Anything V4.5 Pruned Safetensor
Phanes:
Phanes or Protogonus (lit. 'first-born') was the mystic primeval deity of procreation and the generation of new life in Greek Mythology.
This admittedly borderline Cliché name (for something that technically doesnt even create something, but rather just assists in it), was chose without any specific reason, other than me thinking it sounds cool, and there being (so far) no conflicting tags or other models/LoRA's named like that.
This LoRA comes in:
V1: Slightly Overcooked (maybe) which was trained at around 16.8k steps (long story short, i thought epochs would Split/Divide the step count, not Multiply it), still turned out good.
V1.1: Slightly less cooked at 1.6k steps, also slightly more Stable than V1 at higher CFG Values (at least from my testing).
V2: Same as V1.1, only trained with a higher Image Pool size, and slightly different settings, most notably full FP16.
O.K. but what was it trained on?
V1&V1.1: A set of 56 Handpicked Danbooru Images which come from diffrent artists, but have a similar artstyle.
V2: The original 56 + 54 Extra Images.
All Images were Autotagged by Toriatos WD14-Tagger in Automatic1111's Web-UI, and manually cleaned up.
The LoRA itself was trained using Bmaltais Kohya_ss GUI