What I trained before, now I understand, and the model will be modified later

I collected all the emoticons, poses, and illustration pictures in the chronicle and games

I train my eyes using regularization, and this version has a lower number of trainings, resulting in lower eye detail than the previous version

I feel that fan graphics affect the quality of training

@vanilla349 There are definitely multi-player training sets, but they are very few, only a hundred or two, and under the suppression of other single-player training sets, they will not play without input of multi-player prompts

@vanilla349 To be exact, there are only about 700 normal official pictures, and then I put the cropped pictures into it to get about 1,000 pictures. Finally, I put the perfect pictures generated by my ai into it, plus the labeled text, and now there are 2884 files, which is 1442 pictures.

After all, there are so many data sets, even casual training can get better results. I don't even have classification, it is just a simple pot stew.
-----------------------------------------------------
准确来讲正常的官方图片只有七百张左右，然后我将裁剪过后的图片也放进去了 得到了约一千多张图片，最后我将由我ai生成的较为完美的图片也放进去了，加上标记文本现在一共有2884个文件也就是1442张图片。 然后我对图片的裁剪其实没那么严格，毕竟数据集有这么多了，即使是随便训练都能得到较好的结果，我连分类都没有分类，就是简单的大锅炖。

@vanilla349 However, I am currently doing targeted training, that is, classification training, after all, so many data sets, it is not easy to deal with. I am also busy in reality, so only pictures of chocolate are classified.

And this classification training is still in preparation, not officially started.

In my mind, I want to train a perfect lora model, but when it comes to the actual processing of data sets, I always find it troublesome, which is simply cutting, resulting in a lower quality of training than I expected. Moreover, due to the lack of classification in the training set, the controllability of lora is extremely low, and it seems that only the original label text can perfectly control the picture. That is, only I can play the full quality of this lora model more normally.

------------------------------------------------------------
然而我最近正在针对性的训练，也就是分类训练，毕竟那么多数据集，不太好处理。而我在现实中也比较忙，所以说目前分类好的只有巧克力的图片。 而这个分类训练我还在准备中，没正式开始。 我心里想的是训练一个完美lora模型，但是到实际处理数据集的时候，总会嫌麻烦，就是简单的裁切，导致训练出来的质量没有心里预期那么高，并且由于训练集没有分类，导致lora的可控性极低，似乎只有原始的标签文本才能较为完美的控制图片。也就是只有我自己才能较为正常的发挥这个lora模型的全部质量。

@vanilla349 I have a newer version of the lora model in my hand, but after my tests, the difference between them and the old version is very small, almost the difference is the details, such as eye details and some control ability. Personally, I feel that there is not much need to upgrade, because the quality of them is good under my use, but in the hands of the user, because they do not annotate the text, it is difficult to fully play, my training is fully annotate. The training method is low weight merging of lora from the same data set trained by many different large models. It is difficult to play with few prompts.
----------------------------------
我现在手上拥有较为新的版本的lora模型，但是经过我的测试，它们与旧版本的差距很小，几乎就是细节差距，如眼睛细节和一些控制能力。就我个人来感觉就是，没有太大必要进行升级，因为它们的质量在我的使用下，表现挺好的，但是在用户的手上，由于他们没有标注文本，很难完全发挥，我训练时的标注是完全标注。训练方法是很多个不同大模型训练出的多个同样数据集的lora进行低权重合并。在很少提示词下很难发挥。