Type | |
Stats | 2,342 3,345 |
Reviews | (304) |
Published | May 30, 2023 |
Base Model | |
Training | Steps: 1,520 Epochs: 8 |
Trigger Words | arima kana \(oshi no ko\) |
Hash | AutoV2 57CB52B4FD |
[English/中文]
This is my first LoRA model and there had been 7 previous version trained before the fromal version was released. It will be highly appreciated if you could offer me any advice!
Basic Information
Trigger word: arima kana \(oshi no ko\)
Training model: Anything V5 PrtRE (I suggest you use the same model to generate images)
Recommended weight: 0.8-1.0
Example Image Introduction
The relatively extensive datasets enable more elements to be changeable. You'd better look at example images for proper prompts. The example images (originally 600 * 900 definition, 2x SD upsale) are based on Anything V5 PrtRE and AnimeVAE. I don't use any extension except EasyNegativeV2. The same basic parameters will produce the same result.
1-2. high school uniform (Unfortunately, it is hard for AI to learn to the fact that there is a black bowtie on the back of Kana’s head)
3-5. a few sets of clothes appeared in the Oshi No Ko anime (not 100% similar)
6-8. different expressions (I have to admit that the model has quite limited expressions)
9-11. different angles and poses using Controlnet Openpose (When Openpose is used, the result will be less authentic)
12-13. other types of clothes
14-16. greyscale style and other checkpoint models (Obviously, other checkpoint models are less effective than Anything V5 PrtRE)
17. NSFW (What, do you say that Arima Kana only has small breasts?)
Author’s Words (you can skip it)
Arima Kana is a special girl. She has her particular temperament. So I want to produce a LoRA that can paint Arima Kana, not someone who barely has short red hair, bangs and red eyes. It is a quite challengeable task which few present LoRAs (uploaded in CivitAI) can complete. I spent a lot of time painstakingly selecting qualified datasets to achieve this goal.
Acknowledgement
I would be glad if you can like the model. Don’t forget to support me!
Finally but most importantly, Arima Kana is the best!
这是我训练的第一个LoRA模型，在最终版本之前已经迭代过7个版本了......欢迎提出宝贵的意见！
基本信息
触发词：arima kana \(oshi no ko\)
训练底模：Anything V5 PrtRE（我建议也用这个模型生成图片）
推荐权重：0.8-1.0
例图介绍
得益于比较多样的训练集，模型产出的图片也将更具多样性。你可以去参考例图的提示词来实现这些效果。例图原始分辨率为600*900，用了2倍的SD放大，底模Anything V5 PrtRE，使用AnimeVAE，并且除了EasyNegativeV2没有使用任何其他额外拓展。相同的基本参数即可产出与例图一样的图片。
1-2. 高中校服（然而AI其实很难意识到加奈后脑勺还有个蝴蝶结）
3-5. 一些在推子动漫出现过的服饰（没办法完全复刻）
6-8. 不同的表情（我承认这个模型的表情并不是很丰富）
9-11. 用Controlnet Openpose产出的不同角度与姿势（Openpose会导致较低的还原度）
12-13. 其他服饰
14-16. 黑白风格和其它大模型（显然，其他模型的产出效果不如Anything V5 PrtRE）
17. 懂的都懂（什么，你想说帽子没那么大？）
作者的话（可跳过）
有马加奈是一个特别的女孩，有着自己的气质。我希望我做的LoRA可以画出那个特别的有马加奈，而不是随便一个有着红色短发、齐刘海、红瞳的人。这个要求非常难实现，目前的C站上加奈的LoRA都很难做到这点。为此我花了很多时间精心挑选数据集以最大程度还原她的特征。
鸣谢
如果你能喜欢这个模型，那就太好啦，同时也别忘了支持我一下噢~