BananaCatMix

声明：模型禁止任何形式的转载和二次发布，仅供学习参考与经验交流，可自由融合使用，不得用于商业用途或其他非法行为。

封面猫娘是莉喵，路过的大家都可以来ruarua她的头然后点个赞www

The cover cat girl is Leako, and everyone passing by can SQUISH THAT CAT and give her a thumbs up :3

• 中文简介：

（我愿称之为模型大杂烩，边用边融的产物，不小心出了个相对还不错的融模。）

模型基底已无法追溯，目前能记住的包括Cetus、Orange、AnythingV5、PVC、Lyriel、LigneClaire等，可能还有更多，权重值几乎凭脑内配比，少量测试，因此自认为这是“拼凑之作”，但实际使用起来还挺顺手。

该模型是一个偏向二次元画风的融合模型，能稳定输出细节丰富的二次元人物图像，整体画风如示意图所示，喜欢的朋友可以根据需要自行选用。

由于融合了众多奇奇怪怪的模型，该融模对标签（tag）的识别能力相当出色，基本不会太奇怪的词都能识别并准确呈现于画面中，轻松达成你想要的效果。此外，还融合了许多奇妙画风模型，关键词输入正确时还能带来惊喜。同时，由于融合了大量3D设计及写实类风景模型，生成纯风景或物品类图像时也能达到精美效果。

推荐参数：

正向提示词建议使用 masterpiece, best quality 作为起始，反向提示词常用 low quality, worst quality 等，也可参考示例图中的负面提示词，屎山负面提示词欢迎使用，EasyNegative 也欢迎。

CFG 一般使用默认值 7，上下调整其实也都可以，大家可以尝试看看是否有意外惊喜。

采样器方面，测试下来都挺合适，各有优势，个人最常用的是 Euler（出图快）、Euler a（出图快）、DPM++ 2M SDE Karras（画质好）、DDIM（手部崩坏少）。

不启高清修复也能出图，但如果人物或物体太小，细节容易丢失，建议开启高分辨率修复。目前我常用的是 Latent (bicubic antialiased) 和 R-ESRGAN 4x+ Anime6B。前者能生成更多细微细节，但可能出现“乱长”情况；后者则更能稳定保持修复前的图像特征。

v3b 内置了一个 840000 的 VAE，色彩较浓郁，偏好清淡风格的可自行替换。v3e2 版本未内置 VAE，使用时需手动添加，推荐使用 ClearVAE。

• English Introduction：

(I would like to call it a hodgepodge of models, the product of merge while using, and accidentally producing a relatively good merged model.)

The model base is no longer available, I can currently remember Cetus, Orange, AnythingV5, PVC, Lyriel, LigneClaire, etc. There may be more, the weight value is almost from my mind, a small number of tests, so I would call it a piece of ****, but it is quite useful.

The model is a merged model biased towards the ACG painting style, which can stably output the ACG painting style characters with more details, and the overall painting style is shown in the pictures, and you guys can take it as needed.

Relying on a lot of amazing models, this model is still quite good at the recognition of prompts, basically not very strange words can be recognized and reflected in the picture, can easily achieve the effect you want, there are many wonderful styles, will be surprises when the keywords are right. In addition, because many 3D designs and realistic models such as landscapes are merged in, it is possible to achieve beautiful results when generating pure landscape/object-related images.

Recommended parameters:

The positive prompts recommends using "masterpiece, best quality" to start, negative commonly used "low quality, worst quality," etc., or you can use negative prompts from sample images, and EasyNegative is welcome.

CFG I generally use the default 7, up and down can actually be good, you can try to see if there is a wonderful effect.

All the samplers are actually quite suitable, each has its own strengths, self-use is Euler (fast drawing), Euler a (fast drawing), DPM++ 2M SDE Karras (good quality), DDIM (less hand collapse).

If the character or object is too small, the details will collapse, it is recommended to use the Hires.Fix, personally generally used Latent (bicubic antialiased), R-ESRGAN 4x + Anime6B, the former can produce more small details, the latter is better to stabilize the picture characteristics before the Hires.Fix.

v3b version Built-in an 840000 VAE, the color is relatively strong, if you like light, you can adjust to others. And I forgot to add VAE in v3e2, need to set VAE manually when using, recommend ClearVAE.