Sign In

禄良 Luliang | 深空之眼 Aether Gazer(Lora| 差分学习法 Incremental learning | 神童优化器 Prodigy optimizer | Multires noise)

77
681
2
Verified:
SafeTensor
Type
LoRA
Stats
228
Reviews
Published
Jul 8, 2023
Base Model
SD 1.5
Usage Tips
Clip Skip: 2
Trigger Words
luliang
Hash
AutoV2
B5C076F627

V2版本发布,使用了差分学习法和神童优化器,青龙大佬有出视频,感谢社区大佬们的训练方法和算法

V2 Release,using Incremental learning and Prodigy optimizer,Thank you!

V2-1更加激进,可以发挥出大模型本身的特色,使用svd merge直接将人物lora和大眼画风lora通过正负权重融合,相当于只保留的人物特征,其他一切需要靠出图底模发挥,需要稳定的底模才能出图

V2-1 is more aggresive, can have base-model-style-results, using svd merge and need base model which is more stable

V2-2保守一些,使用的是先将大眼的画风lora融进训练底模中,再进行训练,出图会更绑定训练集一些,好处是不容易出古神图

V2-2 can have good results that ignoring the base model you use, and the results are more reliable to the training images


触发词trigger:luliang

权重直接开1,加入1girl可以更稳定点,比较玄学

想换装就调低,大概可以吧,没试

use lower weight if you want to change clothes iguess

想要更还原人设就加对应的prompt,比如half gloves,chinese clothes,round glasses,thigh strap,etc.