Sign In

Style: Konya Karasue

4

Updated: Nov 17, 2026

style

Not recently loaded

Download

1 variant available

bf16 SafeTensor

konya.safetensors

BF16, good balance • 141.83 MB

Verified:

Type
LoRA
Stats

23

0

Reviews
Published

Nov 17, 2026

Base Model
Training
Steps: 1,212Epochs: 14
Hash
AutoV2F439C2D86F
09_greenhouse.png

由deepseek-flash 主导训练
this lora trained by deepseek

这是针对画师 紺屋(pixiv id 10109777)画风训练的风格 LoRA,跑在 MOSS-Image 底模上。

训练集来自其公开作品的精选:冷调海与雪、夜雨霓虹、逆光剪影、无人的教室与站台。不是把某一张图记下来,而是把「怎么打光、怎么留白、笔触停在哪儿」这套习惯学下来。

适合:氛围向插画、单幅叙事、封面与海报、带强烈光线设计的人物图。

风格特征

  • 冷调偏青蓝,暗部不死黑,阴影里带一点蓝绿

  • 逆光与边缘光是主线:头发被光穿透、轮廓发亮、脸常在阴影侧

  • 半厚涂笔触:色块交界不磨平,布料与水面保留笔刷走向

  • 漂浮粒子与景深光斑:空气里有尘、有水汽、有散景

  • 人物常偏小、留白多,情绪靠环境而不是表情特写

推荐用法

提示词写法

触发词 konya 放最前,后面混写「标签 + 一句自然语言」。这个底模的文本编码器是 Qwen3.5,对长句理解好,句子比堆标签更有效:

konya, 1girl, solo, standing, from side, looking at viewer,
school uniform, train platform at night, fluorescent ceiling light,
cool white light, distant city lights, quiet mood

参数

  • LoRA 强度:1.0(首选);0.8 几乎等价,想少一点画风固化时用;0.6 开始掉风格,不建议出图

  • CFG:2.5 – 3.0 起步;要脸部完全打开时提到 5.0

  • 采样器:er_sde + simple,30 步(euler + simple 也可)

  • 分辨率:832×1216 / 1216×832;方形用 1024×1024

  • Skip Layer Guidance(可选):single_layers 22,start 0 / end 0.5

负面词(重要)

low quality, worst quality, normal quality, text, signature, watermark,
artist name, jpeg artifacts, bad anatomy, extra fingers, deformed hands,
blurry, deformed face, asymmetric eyes, fused lips, cross-eyed,
photorealistic, realistic face, adult proportions, long face,
small eyes, narrow eyes, mature, long midface,
overexposed, blown highlights, washed out

最后三个词是必须的:这个模型在正脸近景 + 正面光时极易过曝发白,加上之后肤色才有层次。

三种典型题材的起手式

正脸近景

konya, 1girl, solo, upper body, close-up, face focus,
clear expressive eyes, short midface, small nose, small mouth, soft jaw,
looking at viewer, front lighting, face fully lit, plain background

逆光剪影

konya, 1girl, solo, standing, from side, backlit, rim light, low key,
long hair, wind, dark background, glowing hair edges

环境叙事(人物小、场景大)

konya, 1girl, solo, from behind, white dress, seaside torii,
clear blue sky, calm sea, high key, vast sky, small figure

已知限制

  • 正脸近景会过曝。加反过曝负向词能压住,代价是整体亮度略降

  • 嘴唇和鼻翼极简。这是画师本身的处理方式,但在超大特写下会显得平

  • 远景或小脸时双眼对称性下降,人物占画面 1/4 以下尤其明显

  • 性别漂移。短发 + 深色外套的构图容易生成偏男性的观感,需要 1girl, long hair 明确约束

  • 底模敏感。换底模要重调强度与 CFG,不能直接照搬

English

A style LoRA for the illustrator Konya (pixiv id 10109777), trained on MOSS-Image.

Trigger: konya

Style: cool teal-blue palette, strong backlight and rim light, half-painterly brushwork, floating particles and bokeh, small figures in large quiet environments.

Settings: strength 1.0 (0.8 is nearly identical, 0.6 loses the style), CFG 2.5–3.0 (use 5.0 when the face must read clearly), er_sde + simple, 30 steps, 832×1216 or 1216×832.

Required negatives: include overexposed, blown highlights, washed out — front-lit close-ups blow out without them.

Known limits: front-lit close-ups overexpose; lips and nose bridge are minimal by design; eye symmetry drops on distant faces; short-hair + dark-coat prompts drift masculine.

License: the original artwork belongs to Konya. This LoRA is released for study and non-commercial experimentation. Do not impersonate the artist, sell generated images, or redistribute the weights. It will be taken down on the artist's request.