CivArchive
    Changli (Wuthering Waves) - Krea 2 - v1.0
    NSFW
    Preview 142895066
    Preview 142895268
    Preview 142895308
    Preview 142895305
    Preview 142895315
    Preview 142895323
    Preview 142895328
    Preview 142895334
    Preview 142895336

    Changli from Wuthering Waves, trained for Krea 2.

    Trigger word: changli_zk

    • Krea 2 Turbo (official): weight 1.0

    • DaSiWa CuteDisaster: weight 0.6–0.7

    • 8 steps, CFG 1, er_sde / simple, 832×1216

    The base model matters more than the prompt

    Tested on both with identical prompts and seed:

    • Official Krea 2 Turbo → flat matte cel skin, clean lineart, slimmer proportions, longer hair

    • DaSiWa CuteDisaster → glossy 2.5D face, curvier proportions, shorter bob

    No prompt wording fixed the glossy look. If you want flat anime, change the base model, not the prompt.

    Weight behaviour

    Tested 0.4 → 1.0 across four poses:

    • 0.4 breaks identity — wrong dark hair, red eyes

    • 0.5–1.0 keeps face and outfit stable; higher is sharper

    • On Turbo, eye colour hits canon amber-yellow at 1.0, drifts red-tinted at 0.5

    • Outfit drift comes from prompt wording, not weight

    Checkpoints

    • step 1000 — most flexible with new outfits and scenes

    • step 1500 — sharpest eyes and lineart (my default)

    • step 2000 — slight overfit, exaggerates chest and thighs

    Style prompting

    Krea 2 reads sentences, not tags. One short clause at the end works best (~100 characters):

    Cel-shaded anime illustration, flat colour areas with hard shadow edges, clean linework, painted sky.

    Also good:

    The whole scene as comic-book art: bold ink outlines, flat colour, halftone dots, drawn background.

    Avoid

    • Period style clauses like 1970s cel anime or 1990s cel animation — they replace her outfit with a plain blouse and skirt

    • Style LoRAs stacked on this one — Krea's RetroAnime LoRA turns her dress purple

    • Long style or detail paragraphs — they eat the prompt budget, crop your framing and change her clothes

    • Negative weighting like (glossy skin:-3) — tested, no visible effect at CFG 1

    What it does well

    Salmon-pink to white hair, yellow ringed eyes, white gown over black with the black-and-gold neck piece, red diamond-lattice sleeves and coattail, gold hair ornaments. Holds up in front, back three-quarter, sitting and close-up framing.

    Known limitations

    • No canon long phoenix-ribbon ponytail — you get a short bob with long trailing side locks

    • No feather-shaped burn marks on her left arm

    • Photorealistic prompts fail — the dataset is all anime, so she stays anime

    Training

    musubi-tuner on Krea 2 Raw · 40 images · rank/alpha 32 · lr 1e-4 adamw8bit · 2000 steps on a 4090 · final loss 0.081

    Description

    First release. Trained to 2000 steps, this is the step 1500 checkpoint - sharpest eyes and lineart without the proportion drift that shows up at 2000.

    Use weight 1.0 on official Krea 2 Turbo, or 0.6-0.7 on DaSiWa CuteDisaster.

    LORA
    Krea 2

    Details

    Downloads
    73
    Platform
    CivitAI
    Platform Status
    Available
    Created
    9/16/2026
    Updated
    9/24/2026
    Deleted
    -
    Trigger Words:
    changli_zk

    Files

    changli_zk-step00001500.safetensors

    Mirrors