该模型使用了超过400张高质量的原始图片。
训练图片全部来自该画师的原画集，TV版动画图片素材没有混入其中，得到纯净的矢吹健太朗风格LORA。
只是一个测试模型，风格已经比较相似了，但是角色的细节还有不少问题。待后续再优化吧~
Animagine XL V3版本：https://civarchive.com/models/275386
Pony XL版本：https://civarchive.com/models/347905
Neta Art XL版本：https://civarchive.com/models/463454
SD1.5版本：https://civarchive.com/models/136677
推荐权重：1.0~1.5
--- --- --- --- --- --- --- ---
This model used more than 400 high-quality original pictures for training.
The training pictures are all from the original painting collection of the artist, and the TV version animation picture materials are not mixed in, so as to obtain the pure artist style LORA
It's just a test model. The style is similar, but there are still many problems with the details of the character. I'll optimize it later~
Animagine XL V3 version：https://civarchive.com/models/275386
Pony XL：https://civarchive.com/models/347905
Neta Art XL：https://civarchive.com/models/463454
SD1.5：https://civarchive.com/models/136677
Recommendation Weight：1.0~1.5
Description
FAQ
Comments (2)
I can't believe the generated image quality you post, far detailed and powerful than SDXL I think. Everything is good. Could you please tell me the training resolution and dim you use?
FLUX的精度和细腻程度是远超SDXL的，就是训练不稳定，我还在不断摸索中
训练分辨率和DIM都是常规的1024*1024, 8/4
个人觉得核心还是在打标签上，目前的tag模式有局限，自然语言打标还不够成熟，所以还有路要走