Okay I attempted to train my own "Momo Style" LORA. (Using KOD / NN / Rule34)
It does not have any concepts of characters..
So there is no such thing as "Sabia" or "Neve" in the tagging, if you want those characters you have to describe them.
I did a few quick tests .. with the following rules:
- No Controlnets (meaning the poses are purely prompt based, each image may look slightly different)
- No image touchups or fixups or seed cherry picking
- No Segmenting (each image will have small variations, like clothing changes or background changes)
Known Issues So Far:
I can already see a problem with my tagging, I completely trashed "ponytail" with "folded ponytail" training. I will need to pick a single word tag that doesn't do that.
The image prompts left to right are:
[Describing Sabia as best I can]
[Describing Neve but with short hair and elf ears]
[Describing Neve but with short hair and cat ears]
The background is just "Sunny Day, Village, Huts, Grass"
Feedback appreciated - is the art style over cooked or under cooked?
I will likely need to do another run (or 10..
), so if people think the style is wrong / bad - please say so.
- Focus criticism / feedbacl on the art style (both background & characters), the color palette, line work, shading style, etc.
- Ignore inconsistencies in height / clothing / pose. Those are fixed in a different way.