My English is not good, so I used ChatGPT to translate this introduction

August 8, 2024: Usage Introduction Added

This LoRA has a significant drawback as a facial expression concept—it tends to overfit quite a bit. Since its release, I've occasionally considered re-optimizing the model. However, I found a solution in the meantime: using the ADetailer plugin to individually refine facial details or applying localized img2img inpainting. This isn't an advanced or complex method, but it solves the overfitting issue with this model, so I may not update this model with a new version for now. Thanks to the ADetailer plugin, the LoRA can work as intended.

How to Use It?

First and foremost, the initial step is to ensure that the character has fully closed eyes in the prompt, no matter what. You can input “(closed_eyes:1.4), (sleeping:1.4)”. Regardless of the method you use, just make sure the character's eyes are closed.

Enter the following prompt in the ADetailer plugin (Note: lbw refers to the LoRA Block Weight plugin. If you haven't installed this plugin, simply remove

“:lbw=1,0,0,0,0,0,1,1,1,0,0,0”).

score 9, score 8 up, score 7 up, 
<lora:HJG v10:0.7:lbw=1,0,0,0,0,0,1,1,1,0,0,0>, HJG, (half-closed eyes:1.7),  (sleeping:1.2)

Expand the inpainting feature and set the Inpaint denoising strength range to around 0.45 - 0.55.

In my experience, the final “sleeping” prompt isn’t the only key to influencing the openness of the eyes—"closed eyes" also works well. By following the usage instructions in this document, you should be able to achieve the desired result after a few parameter adjustments. You can also add “light_smile” or any other facial expressions or details. If you're using a character LoRA, make sure to include the character LoRA in ADetailer along with relevant facial prompts.

I am a beginner at training LoRA models, so this LoRA model is a test version.

It is not only suitable for sleeping scenarios; this LoRA's functionality is a facial expression concept that can be used for various purposes. It mainly affects the character's facial features, making the eyebrows and eyes appear particularly relaxed, weak, desolate, and blank. So... you can think of it as your waifu showing a cold expression of utter disappointment towards you (???). To some extent, this LoRA model enables the character to look slightly to the side with empty eyes (thanks to the power of the Pony model).

Trigger words (non-sleeping):

HJG, empty eyes, (half-closed eyes:1.7), (closed eyes:0.9)

Trigger words (lying down and sleeping):

HJG, empty eyes, (half-closed eyes:1.7), (sleeping:1.2)

Optional: You can see all the recommended additional prompt words to use in the LoRA data (dataset training tags) in SD WebUI.

LoRA weight recommendation: 0.5 - 1.1

Please add the following negative prompt words: blush, bags under eyes, lips, eye_contact, short fingers

You can achieve relatively satisfactory expressions by balancing the weights of “half-closed eyes” and “sleeping or closed eyes”.

In non-sleeping scenarios, the weight of “half-closed eyes” will control the openness of the character's eyes.

Please ensure the weight of "half-closed eyes" remains significantly higher than that of "sleeping" or "closed eyes" to prevent the character's eyes from fully closing. The LoRA weight supports fine adjustments to two decimal places, which can be used to challenge the threshold before closure. However, this may lead to errors where the eye color bleeds into the whites of the eyes. The ADetailer plugin might help correct this issue, though it doesn't guarantee a perfect fix.

Reference image↓

①

1girl, solo, long hair, green eyes, black hair, very long hair, <lora:HJG v10:1.01:lbw=1,0,0,0,0,0,1,1,1,0,0,0>, HJG, empty eyes, (half-closed eyes:1.7), (closed eyes:0.9)

The change from (closed eyes:0.9) to (closed eyes:1.43)

②

1girl, solo, long hair, green eyes, black hair, very long hair, <lora:HJG v10:1.01:lbw=1,0,0,0,0,0,1,1,1,0,0,0>, HJG, empty eyes, (half-closed eyes:1.7), (sleeping:0.8)

The change from (sleeping:0.8) to (sleeping:1.3), the girl will also gradually look to the side.

In addition, you can also simply use the prompt: <lora:HJG v10:1.01:lbw=1,0,0,0,0,0,1,1,1,0,0,0>,HJG,empty eyes,(half-closed eyes:1.7). At this point, you can boldly increase the weight of "half-closed eyes". In my own tests, I could increase it up to 2.2 while maintaining normal results. Of course, increasing it to 4.0~6.0 can still generate images, but the eyes are very likely to show errors. I have not tested using ADetailer to fix this error.

Also, in my reference image, I included 'open mouth' so the character would have their mouth open. Well, that's not a big deal.

In sleeping or overhead views of the character lying on the ground, the weight of “sleeping” under the influence of a higher weight for “half-closed eyes” will prevent the character from fully closing their eyes. Increasing the weight of “sleeping” will make the character look more lifeless.

If you have the LoRA Block Weight plugin, these are the parameters suitable for this LoRA: “lbw=1,0,0,0,0,0,1,1,1,0,0,0”
I highly recommend using the LoRA Block Weight plugin because I have little experience in training LoRA models, and there may be issues with image contamination. This plugin will significantly help avoid such problems.

You can view some of the content I posted using this model on PIXIV here: https://www.pixiv.net/users/79440017

These descriptions are just part of my understanding of this LoRA model. Feel free to explore more ways to use it boldly and share with everyone! ❤️

中文精简介绍（由于是精简介绍，如果有什么关于这个模型出现的问题可以可以回复问我）

2024年8月8日 新增使用介绍

这个lora作为一个表情概念缺点很明显，他的过拟合程度比较高。从发布以来我时不时想要重新优化一下这个模型，但我中途找到了解决办法，使用ADetailer插件单独再修复一次面部细节，或是img2img局部重绘，这不是什么高级的、复杂的用法，但解决了这个模型的过拟合问题，所以暂时可能不会更新这个模型的新版本了，lora现在因为ADetailer插件可以正常工作了。

具体的使用方式？

首先第一步，不管怎么样都要在提示词想办法让角色处于完全闭着眼睛的状态，你可以输入“(closed_eyes:1.4)，(sleeping:1.4)”不管您使用什么方式，只要保证角色处于闭着眼睛的状态即可

在ADetailer插件中输入提示词（注意，lbw是LoRA Block Weight插件的内容，如果你没有安装这个插件，就删除“:lbw=,0,0,0,0,0,1,1,1,0,0,0”）

score 9, score 8 up, score 7 up, 
<lora:HJG v10:0.7:lbw=1,0,0,0,0,0,1,1,1,0,0,0>, HJG, (half-closed eyes:1.7),  (sleeping:1.2)  

将重绘功能展开，局部重绘幅度设置在0.45 - 0.55的范围左右即可。

根据我的经验，最后的sleeping提示词不是唯一可以影响眼睛开合大小的关键，“closed eyes” 也同样可行，参考整个文章的使用介绍，你完全可以在几次尝试调整参数后得到想要的结果。你也可以加入“light_smile”或其他各种各样的表情等等一切有关于面部的信息，如果使用了角色lora，也需要把角色lora放入在ADetailer中，包括角色关于面部的提示词。

这是一个不太稳定的测试初版lora模型

功能是影响角色的表情，使其趋近于一种面无表情以及显得很放松的神情状态。

触发词（躺下睡觉）：

HJG, empty eyes, (half-closed eyes:1.7), (sleeping:1.2)

触发词（非睡眠）：

HJG, empty eyes, (half-closed eyes:1.7), (closed eyes:0.9)

重量建议：0.5 - 1.1

负面提示词建议：

blush, bags under eyes, lips, eye_contact

触发词里面，(half-closed eyes)的权重要确保始终高于sleeping与closed eyes，才能保证角色是睁开眼睛的状态。(half-closed eyes)与(sleeping)或(closed eyes)的权重平衡可以灵活微调，有一个小技巧，(half-closed eyes)的权重高于sleeping和closed eyes 0.2的权重值是比较保险的。另外，增加sleeping的权重可以让角色进入更加深度的睡眠姿势，比如(sleeping:1.2)，此时根据技巧的建议，至少要(half-closed eyes:1.4)以此保证不闭上眼睛。

有LoRA Block Weight 插件的话，这是一份适合此 LoRA 的参数：“lbw=1,0,0,0,0,0,1,1,1,0,0,0”

这些描述只是我对这个LoRA模型的部分理解，请大胆探索更多的使用方式，并在评论分享给大家！❤️