Trained exclusively using anime screenshots with multiple characters and scenery. Many different outfits and full body shots are trained just to get the body proportion right.

Recommended Settings:

V1.0 (Illustrious): Trained on Illustrious XL 1.0

Sampler: Euler a

Scheduler: Normal

Steps: 28-30

CFG: 6-7

LoRA Strength: 0.9-1.0

Positive:

absurdres, highres

Negative Prompt:

chromatic aberration, blurry background, blurry, bad quality, worst quality, worst detail, sketch, very displeasing, displeasing, bad ???????, artistic error, lowres, bad hands, signature, artist name, multiple views, artist sign

Base on my testing, using quality tags, like "masterpiece" and "amazing quality", "fix" the character's proportion.

~

V1.0: The LoRA is trained on NoobAI Vpred 1.0. It may or may not work on the Epsilon model.

Sampler: Euler / Euler a

Scheduler: Normal / Simple

Steps: 36-40

CFG: ~5 (Euler) / ~4 (Euler a)

(Based on NoobAI-XL User Manual)

LoRA Strength: 0.9-1.0

Negative Prompt: None, v-prediction model seems to be very sensitive to negative prompt, having it will impact the style a lot, same applies to any artist tags. Use positive prompt if you don't want anything (eg. no headwear, no hair ornament, uncensored).

~

Trigger words:

koi to uso

Some lighting and effects being trained:

light source from back, creates a bright outlines of the character

backlighting

light source from side, usually half of the face is lit

sidelighting

sunset outdoors / candle light indoors

orange light

create a bloom effect with yellow tone

warm colored

chromatic aberration

chromatic aberration