CivArchive
    Artist: 御坂12003 - v1.0
    NSFW
    Preview 6553276
    Preview 6553279
    Preview 6553278
    Preview 6553277
    Preview 6553275
    Preview 6583602
    Preview 6583566
    Preview 6553273
    Preview 6553274

    御坂12003 Style LoRA

    中文 | English

    --

    --

    自备人物模型。主要针对技法进行训练。四肢有问题时降低模型权重。

    Bring your own character model. Mainly for the training of painting techniques. Lower model weights when limbs are problematic.

    --

    仿御坂12003的作画风格的LoRA模型。

    权重建议值:1 (or 0.5~1.2)

    The LoRA model mimics Yuban12003's painting style.

    Recommended weight value: 1 (or 0.5~1.2)

    --

    训练相关/About training:

    训练集共计8张图。双向差异法+多模型蒸馏。

    This version of the model uses bidirectional difference method + multi-model distillation. The training set consists of a total of 8 images.

    loss rate < 1%

    --

    任何问题,欢迎反馈,问题不提出就不会得到解决。

    Any questions are welcome for feedback, as problems will not be solved without being raised.

    Description

    First version.

    FAQ

    Comments (2)

    aertt6950Feb 13, 2024
    CivitAI

    翻了下以前你的lora训练集有上百张几百张图的,这也是我普遍看到的。而现在最近几个lora都个位数图,效果却还可以; 这是得道成仙了吗?

    Chanter
    Author
    Feb 13, 2024· 1 reaction

    算是琢磨出了些门道吧,巧用了下差异法。差异法对训练集的数量要求很少而模型质量很高,双向差异法则在一定程度上提升了差异法的适用范围,多模型蒸馏则能够去除模型的不少杂质,两道工序下来一个不错的模型就基本做好了。不过差异法花的时间比较长,训练步数都在万步以上。这套工序不是万能的,我看情况来。

    LORA
    SD 1.5

    Details

    Downloads
    586
    Platform
    CivitAI
    Platform Status
    Available
    Created
    2/12/2024
    Updated
    9/1/2026
    Deleted
    -

    Files

    style_Yuban12003.safetensors

    Mirrors

    CivitAI (1 mirrors)