kgirl is a semi-realistic East Asian beauty LoRA trained on
ilxl-generated datasets, designed for Krea-2.
This LoRA captures a youthful East Asian beauty aesthetic —
sitting between clean illustration and photorealism. Soft facial
contours, delicate features, smooth skin shading, and a polished
fashion editorial finish define the character's look.
Optimized for fashion, portrait, and editorial-style generation.
Works well across a wide range of scenes — from studio portraits
to outdoor lifestyle and beach resort contexts.
▌Base Model
Trained on Krea-2-Raw — compatible with all Krea-2 variants
(Raw, BF16, FP8, Turbo, NVFP4).
▌Style
Semi-realistic · 2.5D · East Asian beauty · Fashion editorial
▌Recommended Settings
Sampler: Euler Ancestral
Scheduler: Simple
Steps: 20 (2nd pass base)
CFG Scale: 1
Resolution: 768 × 1280
LoRA Strength: 1.00
VAE: qwen_image_vae.safetensors
Clip Skip: 1
Negative Prompt: (none)
Description
kgirl LoRA v1 — Krea-2 / Semi-Realistic East Asian Beauty
Trained on 300 images generated with illustrious XL (ilxl),
featuring a consistent East Asian beauty aesthetic across diverse
beach and resort scenes.
▌Model Details
- Base model : Krea-2-Raw (krea2 arch)
- Network : LoRA rank 16 / alpha 16
- Precision : Full precision (no quantization)
- Steps : 1,500 | Batch size : 2
- Optimizer : AdamW 8bit | LR : 1e-4
- Scheduler : FlowMatch | Timestep : linear
- Resolution : 512 / 768 / 1024
▌Dataset
- 300 images, ilxl-generated semi-realistic style
- Trigger word : kgirl
- Caption dropout : 0.0 (strict caption binding)
▌Character Style
Semi-realistic East Asian beauty — sits between illustration
and photorealism. Inherits ilxl's clean linework and soft
rendering while retaining believable skin texture and lighting.
Youthful facial features, soft facial contours, and a polished
high-end beauty editorial finish. Best results in fashion/
editorial contexts: beach clubs, resort wear, sunset scenes.
▌Trigger
kgirl
▌Recommended Settings
- Sampler : Euler Ancestral | Scheduler : Simple
- 1st pass : 8 steps, CFG 1
- 2nd pass : 20 steps, CFG 1
- LoRA strength : 1.00
- Resolution : 768 × 1280
- Neg prompt : none (ConditioningZeroOut)
FAQ
Comments (1)
Thanks for the lora, but what's going on with the earrings? Why is there such an uncontrolled and chaotic number of them? It really ruins the user experience.