该模型使用了超过200张高质量的原始图片进行训练。
这个模型拥有很好的画风模仿效果
另外，因为是模仿画师画风类型的LORA模型，所以并没有特意针对某些特定角色打tag标签，只有少量代表性角色做了特定训练（隆，春丽 这类）
illustriousXL 版本：https://civarchive.com/models/1456553
Flux版本：https://civarchive.com/models/837136
推荐权重：0.6~1.0
推荐模型：Animagine XL V3
--- --- --- --- --- --- --- ---
The model used more than 200 high-quality original pictures for training.
This model still has a good effect of restoring the artist's style.
In addition, because it is a LORA model that imitates the artist's style of painting, it does not specifically tag some specific characters, and only a small number of representative characters have done specific training (Ryu, Chun-Li, etc.)
illustriousXL version：https://civarchive.com/models/1456553
Flux version：https://civarchive.com/models/837136
Recommendation Weight：0.6~1.0,
Recommendation Model：Animagine XL V3
Description
FAQ
Comments (9)
Was the souce material a shinkiro artbook? Because I keep getting pages with white borders. Other than that great, job, I've been having alot of fun with it.
Yes, some of the original material does come from Shinkiro artbook. And because the original resolution of some materials is relatively low, a mosaic-like effect will appear when generating the image. I will continue to optimize and adjust in the future. -_-
@flyx3 thanks, that explains it. it may be tedious but photoshopping that stuff out before training should fix that. I cant seem to fix it with prompting. adjusting the weight he!ps, but you lose the effect.
This awesome!! if you Did Kinu Nishimura lora it would be Awesome!
为什么我用你这个同样的提示词，底模和VEA，跑出来的图姿势构图相似，但是画风不太一样呢？你这个有点2.5D的感觉，我跑出来的偏2D
可以直接下载放大的原图，丢到SD的图片信息查看器里，可以看到图片的详细原始信息（包括使用的采样器，高清修复参数等等），直接发送到文生图就可以获得完全相同的生成参数
我确实也是这样操作的，可是没有样图好看，负面提示词的embeddings我也都下载了。我传了几张我跑的图，请看看。谢谢
@ivever 森气楼本身画风比较偏2.5D。当然也有不少（例如 饿狼 这类人设）是2D线条为主的。我想问题可能就出在负面词那里，我发现我的文件夹里只剩下一个EasyNegative（负面词是设的模板，懒得改了），所以出图的细微偏差应该就是这里造成的
@flyx3 果然删掉除EasyNegative之外的其他embeddings之后跑图效果就差不多一样了，感谢回复！！