Tutu’s HiSilk (medium fishnet stockings)/图图的嗨丝（中网渔网袜）版本v2.0 (ID: 292421) 综合资源合集综合资源合集

交流Q群：950351015，欢迎来玩，互通有无,

Telegram: https://t.me/+nbU3j7rLEYZkNzVl

I have made fishnet stockings before, but they have never been particularly successful.

The production method has been updated recently, so I recreated it using Difference Extraction Method 3.0+ with real shots.

Although it is still not satisfactory, I think this is the best effect that SD1.5 can achieve.

For this model, a weight of 0.6 can generally achieve better results, and there is no need to make it too large.

For ease of use, all the LORA trigger words I create in the future will be tutututu, four tu.

For this LORA, in order to ensure a high degree of restoration, it is recommended to write in the prompt word:

tutututu, fishnet_pantyhose,

All kinds of large models are very friendly, especially good at two-dimensional.

Instructions:

1. Make sure you use LORA correctly.

2. It is recommended that the number of CLIP termination layers be set to 2.

3. The weight depends on the needs. It can range from 0.4 to 0.9. It is recommended to start with 0.6.

4. The trigger words are: tutututu, fishnet_pantyhose,

5. Don’t add too many negatives, otherwise it will affect the effect of LORA.

6. The prompt word guidance coefficient (CFGScale) can be tried starting from 7.

If you encounter problems during use, you can leave me a message.

Looking forward to your work!

Hope you have a good time.

Thank you for your long-term trust and support.

statement:

1. You shall bear full responsibility for any creative works using this model.

2. You may not use this model to intentionally create or share illegal or harmful output or content, and avoid using this model for malicious, harmful, defamatory, fraudulent or political purposes.

3. If you use this model for commercial purposes, please inform me, thank you.

common problem:

1. What is differential extraction method?

As the name suggests, simply speaking, it is to extract the differences to make LORA. Our general LORA is trained directly with pictures. In this way, the trained LORA will inevitably be contaminated by the base film and other environmental factors, so it will have an unexpected impact on the picture when the picture is produced. The LORA produced by the differential extraction method will remove these contaminations. In this way, even if the weight of the LORA produced is relatively high, the pollution to the picture will be very small, and the model quality will be better. The differential extraction method is more complex to make and requires more steps.

2. Why do you want to do this LORA?

Although there are many similar LORAs, I found that on the one hand, most of these LORAs are not very close to reality. Many of the more popular clothing at the moment are not available, and they are not realistic enough.

In addition, the quality is uneven. Many LORAs have a greater impact on the characters themselves and the graphics, making them inconvenient to use. So I plan to make a series of clothing LORA.

3. Why can’t I take photos like yours?

Everyone's computer configuration and operating environment are different, and the LORA and plug-in versions used, plug-in configurations, etc. will also be different, so please don't pursue the same photos and create your own perfect works!

4. How to get in touch?

If you have any need to customize LORA, or have other cooperation intentions, you can leave a message here directly, or contact Q331506796 (indicate your intention)

5. What optimizations have you made to the difference extraction method? Or are you just talking casually?

A: I have done a lot of optimization on the code involved. There are:

1. Automated model exploration: Automatically generate multiple fusion models from a set of models, similar to the automated process of parameter tuning or model selection, but focused on model fusion.

2. Difference analysis based on Euclidean distance: used to compare the weight differences of two LoRA (Low-RankAdaptation) models.

In dissimilarity analysis, Euclidean distance is used to quantify the difference between the weights of two models. Specifically:

Weight representation: The weight of each LoRA model can be regarded as a point in high-dimensional space. Each element of the weight corresponds to a dimension in the space.

Calculate the difference: By calculating the Euclidean distance between the corresponding weights of the two models, we can get a numerical value that represents the relative position difference of the two weights in the high-dimensional space.

Explain the difference: A larger Euclidean distance means the two models are more different in that weight; a smaller distance means they are more similar.

This difference analysis is used to guide model fusion decisions. For example, if two models differ greatly in a certain weight, this weight may need to be specially processed to ensure that the fused model can effectively combine the characteristics of the two original models.

These optimizations are completely based on my own materials. Because my materials have some special characteristics, they are not difference extraction in the traditional sense. There are some trade-offs.

3. Use the control variable method to adjust each parameter to obtain the best results. For example, I tried 512, 768 and 1024 resolutions, and finally found that 768 worked better.

4. Compare with all other identical models at Station C.

This model has been compared with all other works of the same type at Station C. Although it is still not satisfactory, it is already the best quality among similar models in terms of generalization, pollution, reduction degree, and rendering effect.

图图的嗨丝（中网渔网袜）

交流Q群：950351015，欢迎来玩，互通有无,

电报：https://t.me/+nbU3j7rLEYZkNzVl

之前做过渔网袜，但是一直都不是特别成功,

最近更新了制作方法，所以使用差异提取法3.0+实拍重新制作了一遍,

虽然其实还是不够满意，但是我觉得这已经是SD1.5能达到的最好效果了,

对于这个模型，权重一般0.6就能有比较好的效果了，不需要搞的太大,

为了方便使用，以后我制作的LORA触发词全部都是tutututu,四个tu,

对于这个LORA，为了确保高还原度，建议在提示词中写入：

tutututu, fishnet_pantyhose,

各种大模型都很友好，尤其擅长二次元,

使用方法：

1、确保你正确的使用LORA.

2、CLIP终止层数建议设置为2.

3、权重看需要，从0.4到0.9都可以，建议0.6开始,

4、触发词为：tutututu,fishnet_pantyhose,

5、负面不要加的太多，否则会对LORA的效果产生影响,

6、提示词引导系数(CFGScale)可以从7开始尝试,

如果您在使用中遇到问题，可以给我留言,

期待您的作品！

祝你玩的开心,

感谢您长期以来的信任和支持,

声明：

1、您应对使用此模型的任何创意作品承担全部责任,

2、您不能使用该模型故意制作或分享非法或有害的输出或内容，避免将此模型用于恶意、伤害、诽谤、诈骗或政治用途,

3、如果将本模型用于商业用途，请通知我，谢谢,

常见问题：

1、什么是差异提取法？

顾名思义，简单来说就是将差异提取出来制作LORA,我们一般的LORA都是拿图直接训练，这样训练出的LORA不可避免的会受到底膜以及其他环境因素污染，所以在出图的时候会对画面产生预料之外的影响,而差异提取法制作的LORA会将这些污染去除掉，这样制作出的LORA权重即使开的比较高对画面的污染也很小，模型质量会更好,差异提取法制作起来更复杂，需要的步骤更多,

2、为什么要做这个LORA?

虽然类似的LORA有很多，但是我发现一方面这种LORA大多数不是很贴近现实，很多当下比较流行的服装并没有，而且不够写实,

另外就是质量方面残次不齐，很多LORA对人物本身以及画面的影响比较大,使用起来不够方便,所以打算做一系列的服装LORA,

3、为什么我出不了和你一样的照片？

每个人的电脑配置以及操作环境都不一样，使用的LORA以及插件版本、插件配置等也会有所区别，所以请不要追求一模一样的照片，创作您自己的完美作品吧！

4、如何取得联系？

如果有定制LORA的需求，或者有其他合作意向，可以直接在这里留言，或者联系Q331506796（注明来意）

5、你对差异提取法做了哪些优化？还是只是随便说说？

答：我对涉及代码进行了大量的优化,主要有：

1、自动化模型探索：自动从一组模型中生成多种融合模型，类似于参数调优或模型选择的自动化过程，但专注于模型融合,

2、基于欧氏距离的差异性分析：用于比较两个LoRA（Low-RankAdaptation）模型的权重差异,

在差异性分析中，欧氏距离用于量化两个模型权重之间的差异,具体来说：