ABOUT THE MODEL

Dorothy from MÄR: Märchen Awakens Romance

20251216v2

Trained on NoobAI-XL V pred 2.0

Features

Enhance the smoothness and clarity of the outline. Shift the coloring toward a more cel shading, reducing chaotic mixed colors and excessive light sources. Minimize the greasy or overly glossy look. The tag ‘soft lighting’ can be used to add a subtle glow effect. Hands are currently prone to distortion. All rings were manually removed from the dataset, as this element tends to easily cause deformation and bad hand issues.

Issues

• When the face becomes blurry, consider using ADetailer to fix it.

• When the output becomes blurry, consider using Hire.fx to tix it.

TRIGGER WORDS

Witch Robe

doroshii, pink hair, bangs, twin braids, jewel in braids, witch hat, red eyes, witch robe, puffy sleeves, black fingerless gloves, silver bracelet, white collar, blue brooch, witch boots, 

Maid Costume

doroshii, pink hair, bangs, twin braids, jewel in braids, maid headdress, red eyes,  maid costume, white apron, dark blue bow tie, black round toe heel, black pantyhose,

Cyan Camisole, Denim Shorts

doroshii, pink hair, long hair, bangs, red eyes, cyan camisole with white trim, belt, denim shorts, strappy sandals,

White Collar Shirt, Orange Camisole

doroshii, pink hair, ponytail, bangs, red eyes, white collar shirt, orange camisole underneath, blue trim mini skirt, white shoes, 

Requirements

Requirement

Checkpoints: NoobAI-XL V-pred 1.0

Recommended Tools: Forge

Sample: Euler

CFG: 3~5.5

Recommended weight: 1~0.6

Hires.fix: Recommended Upscale value is around 1.2~2.0, denoising strength is 0.1~0.2.

Recommended size: 768*1344, 832*1216, 896*1156, 1024*1024, 1344*768, 1216*8

32, 1156*896

I’ve set the gallery moderation of this post to “R”, so ???????????????? may not appear on this page.

ARCHIVES

20251212v1

Features

• 4+ outfits

Manually creating the dataset and adjusting the style to match my own preferences took several iterations. The most important lesson I learned early on is that if some images come from AI-upscaled works, their quality is definitely poor. Although these images may appear to have solid color blocks to the human eye, in reality almost every pixel is a slightly different color. In addition, such data often contains various distortions and deformations. If you use this kind of data directly for training without processing it, the resulting model will be extremely bad.

Because of this, I spent a long time cleaning and refining the dataset for this model.