PLEASE READ THIS BEFORE PUTTING THE LORA INTO YOUR FILE!!

在你用这个LORA之前请看这个

Due to I keep experimenting with this dataset, several model's architecture ( if that is the right way to call it ) are different, here is the following model with their own architecture :

由于我一直在试验这个数据集，几个模型的"架构"（如果这是正确的叫法的话）各不相同，以下是各自架构的模型：

26vp1 LOCON

26vp1-2 LOCON

26vp1-3 LOCON

26vp2 LORA

I recommend using 26vp2 or 26vp2-000016
我建议使用 26vp2或者26vp2-000016

Thought this LORA is pretty much a simpler one, with low amount of image dataset and cleaning, it still take ages for experimenting and stuffs. So if you wanted to support me, you can go through ko-fi and buying me a 2$ coffee ☕. I am planning to do models and other style or character LORA in the future.

Simple Description

This model will replicating CanKing, a Taiwanese artist and artist YouTuber's artstyle. It was trained with 26 image of recent CanKing's art.

It was trained with AnyLORA, so clearly it would work the best with AnyLORA. But it would be okay to use it for other anime model, at least I tried it on Counterfeit v3.
https://civitai.com/models/23900/anylora-checkpoint

When using, I suggested using DPM++ 2M Karras or DPM++ SDE Karras for sampling, and on the weightage between 0.6-1. Depends what you wanted. You can see and make up the most suitable weight for you with my plot XY.

And in bunches of generations, I realized 26vp1 family have higher chance of generating... weird stuff and text. All of my image generation do not getting the help of ControlNet, so in theory you could just, slap ControlNet OpenPose and negative prompt (text)

Also, use the CLIP skip 2 and I am not recommending anyone to run kl-f8-anime2 VAE onto 26vp1 family unless you really wanted to. Or else, see the XY plot that contain all of the 26vp1 family with blonde girl smoking cigarettes.

I originally thinking of releasing 16vp3 and 16vp4, but I realized, hell? it is so bad. So I decided to just release the 26 images dataset only.

Any problems with this LORA and stuffs? bang your comment in the comment below.

简单介绍

该模型将复制台湾画家和YouTuber 空罐王的风格。我使用了 26 张空罐王近期的作品图像进行训练。

它是用 AnyLORA 训练的，因此显然用 AnyLORA 效果最好。但我猜也可以用于其他动漫模型，至少我在 Counterfeit v3 上试过。

https://civitai.com/models/23900/anylora-checkpoint

使用的时后，我建议使用 DPM++ 2M Karras 或 DPM++ SDE Karras 进行采样，权重在 0.6-1 之间。您可以通过我的 XY 图查看并确定最适合您的权重。

在成批生成的过程中，我发现 26vp1 家族会出现一些奇怪的东西和文字。我要说我的所有图像生成都没有用到 ControlNet ，因此理论上你可以直接使用 ControlNet OpenPose 和负提示(text)。

此外，请使用 CLIP skip 2，我不建议你在 26vp1 系列上用 kl-f8-anime2 VAE来生成，除非你真的想这么做。否则，请参阅包含所有 26vp1 系列和金发女孩吸烟的 XY 图。

我原本想发布 16vp3 和 16vp4，但我用的过后觉得这。。。太烂的吧。所以我决定还是只发布 26 张图片版的LORA模型。

如果你对 这LORA或者其他的有任何问题或意见，请在下面的评论中发表您的伟大的想法或者问题。