我的英语很狗屎，我使用翻译工具。
我的英語很狗屎，我使用翻譯工具。
英語はクソなので翻訳ツールを使っています。
My English is shit and I use a translation tool.
Weighting:0.5-0.8
==Profile==
This Lora is used to generate a “flesh tool for public toilets”, which uses 6000 images as a training set.
It is mainly used to generate [flesh squat toilet] and [flesh urinal], that is, “stuck in squat toilet or urinal”.
–
这个Lora用于生成“公共厕所的flesh tool”，它使用了6000张图像作为训练集。
它主要用于生成[flesh squat toilet]和[flesh urinal]，也就是“卡在squat toilet或urinal里”。
==Prompt and Negative Prompt==
In practice the [] symbol can be ignored, it is only used for easy identification [Prompt]
[wanaata]:Artist Style.
[flesh squat toilet]: A girl stuck in a squat toilet on the ground, can generate feet, or all limbs stuck in it.
[flesh urinal]: A girl stuck in a urinal on the wall.
[public restroom]、[tiles]、[red background]、[indoors]: Used to describe the background, adding to Negative can adjust the background.
[stuck]、[through wall]: Used to describe the state.
[multiple views]: Only used when the image size is [832×576] or similar, it will generate [pages], usually the left is [frontal view-upper body] and the right is [back view-lower body].
[upper body]、[lower body]、[full body]: Used to control the generated area. Note that “full body” does not generate a complete girl, but a “girl stuck in a squat toilet”. It is also necessary to control which body parts will be stuck under the floor and which will be exposed through the degree of out of frame.
[out of frame]、[feet out of frame]: Mainly used to control [flesh squat toilet].
[frontal view],[top view],[back view],[side view]: Control the viewing angle, you can use [frontal view, top view] or [frontal view, side view] at the same time to generate oblique angles.
–
在实践中，[]符号请忽略不计，它只是用于方便识别[提示]。
[wanaata]:画风。
[flesh squat toilet]:蹲厕里的女孩，可以生成脚底，也可以四肢都卡在里面。
[flesh urinal]:卡在墙上的小便池里的女孩。
[public restroom]、[tiles]、[red background]、[indoors]:用于描述背景，添加在Negative里可以调整背景。
[stuck]、[through wall]:用于描述状态。
[multiple views]:只在图像尺寸为[832×576]或类似时使用，它会生成[分页]，通常情况下左边是[正面视角+上半身(肚脐以上)]，右边是[背面视角+下半身(肚脐以下屁股腿脚)]。
[upper body]、[lower body]、[full body]:用于控制生成的区域。注意”full body”生成的并不是完整的女孩，而是“卡在蹲厕里的女孩”，还需要通过出框程度来控制哪些身体部位会被卡在地板下，哪些会露出来。
[out of frame]、[feet out of frame]:主要用于控制[flesh squat toilet]。
[frontal view],[top view],[back view],[side view]:控制观察角度，你可以同时使用[frontal view, top view]或[frontal view, side view]来生成斜着的视角。
==Size and shape==
It only has three sizes in its training set, so it is suitable to use these resolutions to generate images of different content:
[576×832 and 576×896]: Used to generate individual images of “side”, “top view”, “front view” and other angles.
[832×576]: Used to generate “multiple views” of “front | back” under the premise of “through the wall” and “stuck”, so the generated images are similar to “pages”, with a dividing line in the center of the image.
–
它的训练集中只有三种尺寸，因此它适合使用这些分辨率生成不同内容的图像:
[576×832 and 576×896]:用于生成“侧面”、“俯视”、“正面”等角度的单独图像。
[832×576]:用于生成在“穿墙”和“卡住”前提下的“正面|背面”的“多个视角”，这样生成出的图像类似“分页”，在图像中央会有一条分界线。
=Not relevant to the content and can be ignored=
This Lora is used to generate “flesh tools in public toilets”, using 6000 images as the training set.
In fact, I have prepared a total of 30,000 images as the training set, and spent 15 days manually processing these images. I plan to train a powerful NSFW Lora that can show strong impact.
But starting today, I will have to deal with the death of my relatives for the next half month, and I cannot use my computer. So I can only select part of the training images and train the 0.1 version with fewer [steps].
I look forward to you sharing the images generated using it, Bro, this can give me a lot of motivation!
And take care of yourself, wish you health, no matter where you are from.
–
这个Lora用于生成“公共厕所的flesh tool”，它使用了6000张图像作为训练集。
事实上我一共准备了30000张图像作为训练集，并且花费了15天为这些图像进行人工处理，我计划训练一个功能强大，能表现出强烈冲击力的NSFW Lora。
但今天开始，接下来的半个月要去处理亲戚过世的问题，我无法使用我的电脑。因此我只能挑选一部分训练图像，并且用较少的[steps]训练出0.1版本。
期待你们分享使用它生成的图像，Bro，这能给我很多动力！
以及，照顾好自己，祝你们健康，无论你来自哪里。
=Training parameters, hope this helps you=
It uses SD1.5 as the base model, and the parameters for training are:
它使用SD1.5作为基础模型，训练的参数为:
[batch size:2][epoch:20][Clip skip:1][steps:41370][learning rate:0.8][LR Scheduler:cosine with restarts][LR warmup:0][Optimizer:DAdaptation][Text Encoder learning rate:1][Unet learning rate:1][Dim/Alpha:128][Max resolution:768X768][LR number of cycles:40][Max Token Length:225][Shuffle caption][Noise offset:0.06]