To support my work, you could buy me a coffee

https://bmc.link/cyberdicklang

爱发电 赛博迪克朗 https://afdian.net/a/whmc76

Suggest Prompt: masterpiece,comic

Negative Prompt: low quality,etc.. embedding:unaestheticXLv1,

V3.3 new trainning workflow, new data set, less prompt better quality

—————————————tech upgrade 技术迭代升级—————————————

V1.99 After three iterations totaling 6 times the GPU training time of version 2.0, my graphics card is running hot. This version completely reproduces the classical Japanese anime style, and has good style stability. The problem is that the action is not easy to control, this is a good version if it is just a portrait.

经过三次迭代总计6倍于2.0版本GPU训练时间，我的显卡已经要发烧了。这个版本完全复现了古典日式动漫风格，并且有良好的画风稳定性。问题是动作不太好控制，单纯出人像的话这是一个不错的版本。

v2.0 Iterated training methods, larger footage, better performance 迭代了训练方法，更大的素材集，更好的表现

Used 500 high-quality images for training, stronger painting style fixation, better face, simpler prompt requirements.Unless you use very strong SDXL style words, the performance is usually very stable.

使用了500张高质量图片进行训练，更强的画风固定，更好的脸部，更简单的prompt要求。除非你使用非常强的SDXL风格词，通常情况下表现非常稳定。

—————————————tech upgrade 技术迭代升级—————————————

v1.3 more better on both webui/comfyui, colorful and retro anime style

screen has richer color performance than the 1.2 version, and can better maintain the animation style. Even using webui can generate good images. Of course, if you want to get the best experience, it is recommended to use comfyUI and embedding: unaestheticXL.

画面较1.2版本获得更丰富的色彩表现，且能够比较好的维持动漫风格，即使使用webui也可以生成不错的图像，当然如果想获得最好的体验还是建议使用comfyUI和embedding:unaestheticXL

v1.2 better ANIME style and DARK THEME 更好的动画风格和暗色调表现

uses the two-dimensional base mode to retrain to obtain better painting style stability, and no longer requires sdxl-specific embedding (of course, it will be better to add it). If you want richer colors, it is recommended to add negative cue words to the v1.0 version. The next version will enrich the color performance on the basis of stable painting style.

使用了二次元底模重新训练获得了比较好的画风稳定性，不再强制需要sdxl专用embedding（当然加了会更好）。如果想要更加丰富的色彩还是推荐v1.0版本加负面提示词。下个版本会在稳定的画风基础上丰富色彩表现。

v1.0 recommend use on ComfyUI with embedding:unaestheticXLv13 (webui not supported yet)

this version recommends using negative hints for better 2D effects: unaestheticXLv13, unaestheticXLv1, 3D, cartoon, jpeg artifacts, signature, watermark, username,

这个版本推荐使用负面提示 实现更好的2D效果： unaestheticXLv13,unaestheticXLv1, 3D, cartoon, jpeg artifacts,signature,watermark,username,

—————————————闲聊—————————————

The reason for the creation is that SDXL currently does not have an anime model with a darker theme. It is really uncomfortable for people who are used to darksushi, so they can only try to make one by themselves.

The current problem is that not using the cue word tool will leave too much 3D texture, as well as a not-so-good face shape, which can be adjusted using the taboo double seal.

Using unaestheticXLv13 and unaestheticXLv1 at the same time can restore the 2D painting style well

I used 300 super retro bells and whistles for training, mainly because I used it for a long time to debug the parameters, and I had a headache in the direction of sifting images. At present, the training feeling of SDXL is quite different from before. Hope to give more comments, the model will continue to be updated.

The following improvements are expected:

More retro character style

Better control

Try to upgrade the version in two ways: train/merge

suggest negative prompt: unaestheticXLv13,unaestheticXLv1 （Yes, two at the same time

创作理由就是SDXL目前没有一个比较暗色调主题的动漫模型，对于用惯了darksushi得人来说真的是难受，只能自己尝试做一个。

目前的问题是不使用提示词工具会残留过多的3D质感，以及不太好的脸型，可以使用禁忌的双重封印进行调整，

同时使用 unaestheticXLv13和unaestheticXLv1,可以很好的还原2D画风

使用了300张超级复古花里胡哨的素材进行训练，主要是调试参数用了好久，还有筛图的方向有点头疼 。目前SDXL的训练感觉和之前区别还是挺大的。希望多多提意见，该模型会持续更新。

预计会有以下改进：

更加复古的人物画风

去除3D质感

更好的控制性

尝试以train/merge两种方式进行版本升级

推荐使用负面提示： unaestheticXLv13,unaestheticXLv1 同时使用

Contact:

B站小红书 赛博迪克朗 欢迎点个关注吼

Email: [email protected]