[LoRA] K-AIgirl-face
模型简介:也算兑现承诺了吧,把我早期训练并使用的lora分享给大家。
基于schoolmax 2.5d训练,后来我对比了一下这个模型和0.5chillout+0.5AOM2的完美世界基本上区别很小,
所以可以认为是基于chill训练的,但是schoolmax 2.5d颜色比较舒服一点。
这个模型对于脸部权重的修正比较中性(但拉高权重还是改变挺大的,所以不是欠拟合),因为我不是拿特定脸训练的,而是从生成结果里选择了一些自己觉得好看的图片进行二次训练。
样本数量从70到1300都有,但是1300的大样本(K-AIgirl-face)训练不是很成功,反而是低分辨率小样本的K-AIgirl-face2效果比较好,K-AIgirl-face3则有比较大的偏色问题,过拟合情况比较严重。
推荐用法:
0.2 K-AIgirl-face-re/K-AIgirl-face2
+
0.2 koreanDollLikeness_v15
作为基底,配合chilloutmix,可以大大改善生成人物的脸型和脸部背光的问题,对于其他lora的兼容也很好。
甚至可以把这两个lora融到大模型里面,毕竟lora用多了速度损失实在太大。
~~~~~~~~~~~~~~~~~~~~~~~o(* ̄︶ ̄*)o~~~~~~~~~~~~~~~~~~~~~~~
示例提示词和参数:
Prompt: masterpiece,extremely detailed CG unity 8k wallpaper, realistic,((sfw))
//
1girl,white hair,wavy hair,long hair,
//
(see-through wedding dress:1),no panties,(wedding,bride:1.2),in church,standing
//
Negative prompt: (worst quality, low quality:1.4), (bad anatomy), watermark, signature, text, logo,contact, (extra limbs),(3d),loli,Six fingers,Low quality fingers,EasyNegative:1,
(((missing arms))),(((missing legs))), (((extra arms))),(((extra legs))),less fingers,lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry,
panties:1.5,penis:1.5,
Steps: 28, CFG scale: 7.5, Sampler: DPM++ 2M Karras, Seed: 1463833476, ENSD: 31337, Size: 640x960, Model: 0.5(chilloutmix+Yozora out24) + 0.5(++0.85(0(camelliamix25D_v10) + 1(hands32)) + 0.15(chilloutmix_Ni-fp16)-fp16-no-ema), Model hash: 9497fc2694, Hires steps: 10, Hires upscale: 2, AddNet Enabled: True, AddNet Model 1: K-AIgirl-face-re(2a708595bea4), AddNet Model 4: koreanDollLikeness_v15(3c8b9d1e9b98), Hires upscaler: ESRGAN_4x, AddNet Module 1: LoRA, AddNet Module 4: LoRA, ControlNet Model: control_sd15_openpose , AddNet Weight A 1: 0.4, AddNet Weight A 4: 0.2, AddNet Weight B 1: 0.4, AddNet Weight B 4: 0.2, ControlNet Module: none, ControlNet Weight: 1, ControlNet Enabled: True, Denoising strength: 0.5, ControlNet Guidance End: 1, ControlNet Guidance Start: 0, Clip skip: 2
页:
[1]