The LoRa trained based on NIJI6 images has a stronger NIJI style, better effects on faces and limbs, and can be used with other LoRa.
You can easily generate a mature muscular man.
You can use prompt words to adjust the clothing you want to generate, including shoulder armor, wrist guards, belts and other elements, with a strong Chinese style clothing.
I recommend using 0.4- 0.6
most Sampler are good, like Euler a, Dpm 2m/3m Sed...
for Schedule type, I think Karras is the best, but you can try othes (maybe?)
About V4
I have more training for the face, I do beelive you don't need facial repair
Clothing will be better removed and replaced, you can use nude or topless prompt to control
Feedback is welcome
Description
FAQ
Comments (7)
这个lora生成的任务真的很好,盔甲也很霸气。但我感觉对画面的污染有点重了,有点过拟合,生成图片的背景总是没有细节,可能需要对数据源进行更精细的打标,或者使用正则化,以提升lora的泛化性(表现力)。如果你追求的就是这种感觉,那当我没说^_^,整体还是很棒的。
感谢反馈,我无法找到盔甲细节,背景和过拟的平衡点,或许加入一些远景图丰富背景会好一些?或者将原素材重绘放大一下可能会好一点?我还没有使用过正则化,需要学习一下哈哈
@Youqu 以我的经验来看,就是打标不精细,我看你模型用了350张素材,打标的词组总共不到一百个,而且训练的是风格模型,那大概率会过拟合…一般350张素材的风格lora,打标的词组都大几百上千个了。。。这样模型就会分不清什么是什么。至于正则化,会让模型的触发变得困难,并且大大提高你的工作量,感觉用的人不多,反正我是不用的,我懒。。。
当降低lora但强度,看似对画面的污染变好了,但盔甲以及人物的控制力同样也会变差。该说不说,你的素材质量真好,过拟合出来的图片也挺好的。。。
@Xabsurd 我也不喜欢正则化lol
@Youqu 求教一下,是不是你用的素材各种尺寸的都有。感觉和我碰到的困境差不多……
@akashic921 我基本都是统一的尺寸,1024*1024偏多,除非要全身或者裤子的素材我才会拉长画面
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.


