illustrious LoRA super highleg

详情

模型描述

v3.6

基于 Illustrious-XL-v1.0 进行训练。

Illustrious-XL-v2.0 版本请见 Hugging Face样本生成太麻烦了

说明

这是一个根据示例图像强化高开衩角度的 LoRA。

主要训练了以下三类服装:

  • 连体紧身衣 (leotard)

  • 连体泳衣 (one-piece swimsuit)

  • 水手连体紧身衣 (sailor leotard)

虽然仅训练了这三种,但也可应用于其他服装。

使用方法

在提示词中加入 super highleg,并将 LoRA 权重调整至 0.5~1.25 左右。

加入 midriff, stomach, navel 可使类似短裤的效果更自然。

虽然未明确训练正后方视角,但若加入 thong 等相关词,可能会获得较理想的效果。

当生成类似裤带的部件时,可在负向提示中加入以下内容以抑制:

  • panties

  • panty straps


以下为截至 v2.0 的说明

为测试 CAME 优化器而训练。虽已有相同概念的模型,但作为实验及个人用途仍进行了训练。

eps 版本示例图像模型:NONAMEmix_v11(1.40 之前)

训练模型:AnyIllustrious-XLV1.1

2024/12/10 添加 v_prediction 版本

2025/01/17 仅添加 v_prediction 版本的 v1.40

2025/04/18 添加 1.50-eps(使用 AnyIllustrious-XL-v1.1.0 在 1536 分辨率下训练)

2025/04/21 添加 2.0-eps(使用 Illustrious-XL-v2.0 在 1536 分辨率下训练)

所有版本补充说明:加入 "wedgie" 可使裆部明显紧贴。


v2.0

基本与之前的版本无异,请参考示例图像。

It is basically the same as the previous version. See sample images.


1.50-eps

全面重新调整了超参数,以改善过拟合……应该是这样。

数据集完全未变,因此使用方式与之前一致。

Hyperparameters were completely revised to improve over-learning...supposedly.

The data set has not changed at all, so the usage remains the same as before.


v1.30-any-lbw & v1.30/1.40-anynoob_v_pred-lbw:

整合并训练了总体数据集,其中包含约 30 张使用 v1.13 生成的图像,使剪裁线更为清晰。此外,由于测试了比 v1.13 更多样化的模型,因此应能适配多种插画风格。

权重:1

已按以下方式调整块结构:

1,0,0,0,1,1,1,1,0,0,0,0

标签提示:

  • super highleg gym uniform, :将标签不以逗号分隔而直接连写,有时可获得更有效的服装效果

  • 提高触发词权重会有轻微效果

    • (super highleg:1.4)
  • 当在裙子等衣物下添加内搭时,请使用以下标签

    • leotard under clothes, leotard under pants,
  • 若在水手服风格中,下胸部位出现不明部件,可尝试在负向提示中添加下列标签,可能使其消失(可能受《舰娘》等装备影响)

    • armor,

English:by DeepL.

We have organized and trained the overall data set, including about 30 images generated using v1.13.This should result in a clearer cut line.We also tested with a more diverse model than v1.13, so it should be compatible with a variety of illustration styles.

Weight: 1

The blocks are adjusted as follows

1,0,0,0,1,1,1,1,0,0,0,0

Tagged TIPS:.

  • super highleg gym uniform, :Some outfits are more effective when the tags are joined together and not separated by commas, such as

  • Raising the trigger word weight has a slight effect.

    • (super highleg:1.4)
  • If you are putting something on under a skirt or other garment, please use the following tags

    • leotard under clothes, leotard under pants,
  • If you see mysterious parts around the lower breast in the sailor suit system, add the following tag to the negative prompt and it may disappear

    • armor,

v1.13-last:

所有示例图像均通过 Kohya HRFix 从头以高分辨率生成。

为避免过度偏向特定风格而进行了调整,因此效果可能略弱。

权重:1~2

※权重过高会导致其他影响增强,请根据 seed 值、模型和提示词适当调整。

在原版 Illustrious 或相近模型中,以下设置也可能有效:

lbw=0,1,1,1,1,1,1,0,0,0,0,0

已验证的服装类型:

  • 连体泳衣(竞技泳衣、校服泳衣)

  • 比基尼

  • 连体紧身衣

  • 连衣裙

  • 非传统巫女服

  • 日式短裤(buruma)

  • 连体衣

  • 体操服

其他服装类型也应能实现高开衩效果。

English:by DeepL.

Learned for testing the CAME optimizer.There is already a model with the same concept, but I trained it as an experiment and for my own use.

All sample images are generated in Kohya HRFix at high resolution from the beginning.

I have tried to adjust the images so that they do not lean too much toward a particular style, so they may not work as well as they could.

sample image model:NONAMEmix_v11

Learning model: AnyIllustrious-XLV1.1

Weight: 1~2

*If the weight is raised too high, other effects will increase, so change the weight according to the seed value, model and prompts.

Confirmed attire:

  • one-piece swimsuit (competition swimsuit, school swimsuit)

  • bikini

  • leotard

  • dress

  • nontraditional miko

  • buruma

  • bodysuit

  • gym uniform

Perhaps other outfits should be highleg as well.

此模型生成的图像

未找到图像。