Zenless Zone Zero NPC Girl Style

Trained exclusively using my own screenshots taken by the IN-GAME CAMERA.

Update:

• Dataset 2.0: Added ~80 unique images (360 total) to help seperating the character from the style more, as well as adding more NPC outfits (~20 total, all trained pretty well) and scenery to increase model capability and background variety.

• Dataset 1.1: Added more images for the two girls (first two images in 1.1 (Anima) showcase) who was not well trained in 1.0 dataset.

• Added V1.0 (Lite): Only keep necessary blocks for the style. This fixes detail loss (especially for hair style and clothing) and body proportions for some characters.

Recommended Settings:

V1.1 & 2.0 (Anima): Trained on Anima base-v1.0

Sampler: ER SDE or Euler*

Scheduler: Beta or Simple*

Steps: 30 or 12*

CFG: 4-5 or 1*

LoRA Strength: 0.8-1.0

(*) with Turbo LoRA

~

Triggers:

Character:

npc girl, zenless zone zero

npc girl \(zenless zone zero\)

or natural language saying something like

A npc girl from zenless zone zero

Style (Use natural language for 2.0):

zenless zone zero \(style\), 3d

or natural language saying something like

A zenless zone zero style 3d render

~

Extra Tip: since Turbo LoRA with CFG = 1 can't take any negative prompt, use

no hair ornament, no hairclip, no hair bow

to remove unwanted hair accessories or just use a different seed.

~

V1.0 & 2.0 (Illustrious): Trained on Illustrious XL 1.0

Sampler: Euler a

Scheduler: Normal

Steps: 30

CFG: 6-7

LoRA Strength: 0.9-1.0

~

Trigger words:

Character Tag:

zzznpc girl

Style Tag:

3d

For 2.0, use:

zzzstyle

~

Random Character Generator (Outdated, but still works):

If you have Dynamic Prompts, or any kind of wildcard nodes/plugins, here is an one liner to generate a random character:

{closed mouth|parted lips}, {jitome|tareme|tsurime}, {brown|grey|yellow|pink|blue|red|aqua} eyes, {upper body|cowboy shot, {black|red|brown|pink|blue} skirt, pleated skirt, {plaid skirt|}}, {{blue jacket, open jacket|aran sweater, red sweater, v-neck|yellow cardigan, v-neck|}, red bowtie, {plaid bowtie|striped bowtie|}, white shirt, collared shirt, school uniform|{black|grey|blue} hoodie, hood down, {black|grey} hood, drawstring, clothes writing, {{black|grey} elbow patches|}, casual}, {brown|blonde|green|orange|blue|black|purple|grey|aqua} hair, {long hair, {twintails|ponytail|curly hair}|short hair, {bob cut|flipped sidelocks|}|medium hair, {wavy hair|flipped hair|low twin braids}}, {{short|swept|blunt|double-parted|arched} bangs|bangs pinned back|middle part, parted hair, forehead}, {medium breasts|small breasts|flat chest}, looking {ahead|to the side|at viewer}, {from side|from above|from below|}, {indoors|outdoors, {day|night|dawn|dusk|}}, {sidelighting|backlighting|}

Note: Because some outfits are inconsistant and barely trained, I didn't put them in here. The main focus of the LoRA is the style.

~

Time of Day (light color, only for outdoors):

dawn | day | dusk | night

Light Direction (affect shading on character's face, can also put both in negative prompt to get front lighting):

sidelighting | backlighting

Shadow (more likely to put character under direct sunlight):

shadow

Shade (around character or in the shade):

shade

Reflections (these are introduced for training after V1.1):

reflective floor, reflective liquid, reflective surface, reflection

Some useful negatives that may improve the results:

low poly, model clipping, bad perspective, aliasing