Usage: Recommended strength between 0.8 and 1.0.
Trigger/Prompt: "slaytheprincess, cross-hatching , greyscale, messy lineart " at the start of your prompt. Weights do work with anima, but you need much higher value than SDXL based models, so using (messy lineart:1.8) is perfectly fine if you want a stronger style.
Tested on and in order of style strength:
animapreview3_base (CFG 5.5, 30 steps, Euler)
AnimaYume v0.4 (CFG 5.5, 30 steps, Euler)
Anima Turbo LoRA v0.1 with both Animayume and animapreview_v3 (CFG 1, 12 steps)
RDBT Anima (CFG 2, 12 steps, Euler), style is rather weak.
Know issues:
clothes often transform into the princess' dress unless precisely prompted.
a crown often appears on characters.
severe character bleed, I'll try to fix later.
Description
Comments (4)
with which prompt do you invoke different paths (different princeses)? Is there all the characters or just a few?
This is purely a style LoRa, not character. The Princess appearing in images is a side effect of her being so ubiquitous in training data unfortunately.
I do intend to make a character lora including most different Princesses with a different tag for each though, I just need to find a way to avoid extreme character bleed because of how similar they all look (flowing clothes, long hair, etc).
@ynead thats nice. I think by just tagging training data with proper princess it would give way for style lora to be "all character" lora. Which is not perfect but still.
@Ses_AI Yes but actually no, I already tried that approach. You automatically end up with concept bleeding accros princesses variants, such as everyone getting a tail + horns pretty often, Spectre princess never rarely being white AND translucent and so on. Only radically different variants such as Nightmare, who is wearing full black + a mask come out fine.
Maybe my training pipeline or my dataset are to blame though. Currently the best I can do is a lora with ~3 variants before it starts to get messy. We'll see.