看这边：https://civitai.com/models/212446/bowser-rework

See also: https://civitai.com/models/212446/bowser-rework

我知道这个网站上已经有了我制作的Bowser大模型和另一个人训练的LoRA模型，但是实话说，我之前制作的那个大模型确实存在着过拟合的问题，它污染了其他的prompts，导致即使你不输入Bowser，结果也总是会带有Bowser的某些特征。而另一位仁兄@RockRockRock训练的Bowser LoRA模型，生成Bowser图像的质量不高，而且泛化性不足，在不同的模型下，生成的质量参差不齐。因此，我自己训练了一个Bowser LoRA模型，它具有如下特点：

1. 仅需要使用Bowser作为触发词，易调用性很不错

2. 还原效果优秀，同时没有发生画风过拟合的状况，因此在不同的基底模型上使用的时候，画风是跟着基底模型的，同时角色特征保持不变

3. 可以使用eyes_closed，arms_akimbo，hands_open，clenched_fists，running，jumping，breathing_fire，crawling等prompts来控制Bowser的动作

4. 可以使用back_view，left_view，front_left_view，top_view等词来控制视角

5. 可以加上SSBU来使用任天堂明星大乱斗特别版的画风

总之，这个模型在易调用性、泛化性、还原性方面的效果都很优秀。

这一次我从自己的Nintendo Switch里的马力欧派对中截取了76张图，从任天堂明星大乱斗特别版截取了16张图，总共92张图，手动抠图并替换为白底之后，手动标注tags，制成了训练集。训练了30个epoch，重复次数为6。

实例图片使用了novel ai模型和Protogen v5.3模型生成。

其他的非Bowser图片是用来说明模型没有过拟合，不会出现污染其他prompts的情况的。

English translated by Bing Chat:

I know that there are already a large Bowser model that I made and another LoRA model trained by someone else on this website, but to be honest, the large model I made before had overfitting problems, it polluted other prompts, resulting in even if you don’t input Bowser, the result always has some features of Bowser. And the Bowser LoRA model trained by another gentleman (@RockRockRock), the quality of generating Bowser images is not high, and the generalization is insufficient, the quality of generation varies under different models. Therefore, I trained a Bowser LoRA model myself, which has the following characteristics:

1. Only need to use Bowser as a trigger word, easy to call

2. Excellent restoration effect, while not overfitting the style, so when using different base models, the style follows the base model, while the character features remain unchanged

3. Can use eyes_closed, arms_akimbo, hands_open, clenched_fists, running, jumping, breathing_fire, crawling and other prompts to control Bowser’s actions

4. Can use back_view, left_view, front_left_view, top_view and other words to control the perspective

5. Can add SSBU to use the style of Super Smash Bros. Ultimate

In short, this model has excellent results in terms of ease of use, generalization and restoration.

This time I took 76 pictures from my own Nintendo Switch’s Mario Party and 16 pictures from Super Smash Bros. Ultimate, for a total of 92 pictures. After manually cropping and replacing them with white backgrounds, I manually labeled tags and made a training set. Trained for 30 epochs with a repetition of 6.

Novelai_final and Protogen v5.3 are used to generate example images.

The other non-Bowser images are used to illustrate that the model does not overfit and will not contaminate other prompts.