LoRA-Nekopara猫娘乐园-vanilla香子兰版本nekopara-vanilla-v5 (ID: 128934)

Recommended versions :"vanilla-newlora-ver1","super-vanilla-v4","nekopara-vanilla-v5", the rest are not recommended

vanilla-newlora-ver1

This version supports multiple people

But it will be affected by Hires. fix

The most popular were chocolate and vanilla

It is forbidden to sell the images generated by this model for profit, and it is forbidden to sell this model. Thank you
禁止将此模型生成出的图片进行出售盈利行为，禁止出售此模型,谢谢
Please watch the release introduction for detailed information
请观看版本介绍 有详细的信息

The model generalizes and transfers very well, and you can use it on almost any submodel (except SDXL).
If you want to restore the original painting style, you need a lot of prompt words to correct, the model's data set contains all nekopara images, the game illustrator artbook and fanclub cover, theoretically it can restore all images

---

vanilla-newlora-ver1

This version is trained for classification

Older versions of the dataset are discarded

Stronger support for multiple people

Supports the human hand model

The overall painting style shifts towards nekopara4

More rounds of training were performed on high-quality datasets

High-quality data sets trigger the word vanilla

Generalization trigger word xiangcao

Hand trigger word shouban

r18 Trigger word nsfw

Multi-person trigger word girls

---

描述:

Recommended weights 0.6-1.1

This release fixes the blurred eye issue, but the issue persists? I don't know why? The blurred eyes are solved, but they don't look like the eyes in the original painting?

It doesn't add new data to the original, and it may be better in generalization than the previous version, maybe, well, I don't know exactly how it is.

And then, about this model

The model has been trained in an astonishing 500,000 steps

Why is this, because of my data volume and my training method, my data volume without regularization atlas has more than 1300 sheets, and the number of training steps in a round has 1.3w steps, but the training method of my model is multi-model fusion, so it is necessary to train multiple LoRa with different base modes, resulting in crazy superposition of training steps.

Finally, this model will not be updated in a short period of time, because it is close to the original painting, and I am going to train other cat women, such as coconut azuki..

推荐权重 0.6-1.1

此版本修复了眼部模糊问题，但该问题依然存在？我不知道为什么？眼睛模糊解决了，但是不像原画的眼睛了？

并没有在原有的基础上增加新的数据，在泛化性上可能会比上个版本更好，可能，嗯，我也不知道它到底如何。

然后就是，关于这个模型

这个模型的训练步数达到了惊人的500000次以上

为什么会这样呢，因为我的数据量和我的训练方法，我的数据量不加正则化图集就有1300多张，一轮的训练步数有1.3w步，但是我的模型的训练方法是多模型融合，所以需要多次训练多个不同底模的lora，导致训练步数疯狂的叠加。

最后就是，这个模型短时间内不进行更新了，因为它已经接近了原画了，我要去训练其他的猫娘了，比如椰子红豆...

训练词语: xiangcao,blue eyes