CivArchive
    Art style:xiaoyue style(个人风格) - v1.0
    Preview 148319582
    Preview 148319711
    Preview 148319586
    Preview 148319713
    Preview 148319628
    Preview 148319622
    Preview 148319626

    逆光通透二次元画风

    建议使用MOSS-Image-heal

    细线条、干净的皮肤上色、头发上的细碎高光和脸上的小块光斑,适合逆光、波光粼粼的海面、华丽服装和双人贴脸这类画面。

    推荐设置

    • 触发词:xiaoyue(中文提示词用 xiaoyue画风),放在提示词最前面

    • LoRA 强度:1.0(测试都用的 1.0)

    • 底模:MOSS-Image-heal

    • 采样:euler / simple,50 步,CFG 3,shift 5

    • 建议加跳层引导 SkipLayerGuidanceDiTSimple:single 22,0~50%

    • 尺寸:横图 1536×1024 测试最多

    • 负面:worst quality, bad anatomy, watermark, jpeg artifact, blurry

    示例提示词

    xiaoyue. Two girls leaning their heads together on a sunny balcony. The girl on the left has short black hair and amber eyes and makes a peace sign with a wide smile. The girl on the right has long silver hair with a gold hairpin and violet eyes and looks at the viewer with a gentle smile, wearing an ornate white and gold dress. Bright sky with large white clouds, soft light flares.

    英文标签、英文自然句、中文标签、中文自然句都参与了训练,四种写法都可以用。

    训练信息

    • 底模:MOSS-Image-heal,用 KohakuMoss 训练

    • 数据:58 张,其中 37 张为原图,21 张是用 v1 生成后人工挑选的补图

    • 标注:PixAI 标签 + GBC 自然句,人工逐张核对,四种 caption 混训

    • rank 32 / alpha 128,学习率 1e-4 恒定,RTX 5090 每步 2 张

    • 本版本为第 3400 步,每张图约学习 117 遍

    已知限制

    • SFT 自带的光影比较强,暗场景里仍会带一些 SFT 的电影感暖光

    • 大面积高亮的场景容易偏白、发淡,提示词里少写过曝,多写深色的物件和阴影会好一些

    • 眼睛的虹膜细节比原画简单,脸也偏圆一些

    • 这一档放在 heal 上用时颜色会偏灰偏淡,建议在 SFT 上使用


    Sunlit Anime Glow Style

    Thin linework, clean skin shading, fine highlights in the hair and small patches of light on the face. Works best with backlighting, glittering water, ornate costumes and close two-character shots.

    • Trigger: xiaoyue (xiaoyue画风 for Chinese prompts), at the start of the prompt

    • LoRA strength: 1.0 (all tests used 1.0)

    • Base: MOSS-Image-heal

    • Sampler: euler / simple, 50 steps, CFG 3, shift 5

    • Skip layer guidance recommended (SkipLayerGuidanceDiTSimple): single 22, 0–50%

    • Size: mostly tested at 1536×1024 landscape

    • Negative: worst quality, bad anatomy, watermark, jpeg artifact, blurry

    Training

    • Trained on MOSS-Image-heal with KohakuMoss

    • 58 images: 37 original images and 21 images generated with v1 and hand-picked

    • Captions: PixAI tags and GBC natural language, checked by hand, in English and Chinese

    • Rank 32 / alpha 128, constant LR 1e-4, 2 images per step on an RTX 5090

    • This version is step 3400, about 117 passes per image

    Known limits

    • SFT's own lighting stays fairly strong, and dark scenes keep some of its warm cinematic look

    • Very high-key scenes can turn pale. Fewer overexposure terms and more dark objects and shadows help

    • Eyes have simpler irises than the source art, and faces lean rounder

    • On heal this checkpoint turns greyer and paler, so SFT is recommended

    Description

    LORA
    MOSS Image

    Details

    Downloads
    18
    Platform
    CivitAI
    Platform Status
    Available
    Created
    11/8/2026
    Updated
    11/8/2026
    Deleted
    -

    Files

    xiaoyue_v2_heal_s3400.safetensors

    Mirrors