chtholly-lora / README.md
Wang
Update README.md
6137644
|
raw
history blame
2.68 kB
metadata
license: openrail
datasets:
  - Carzit/SukaSuka-image-dataset

chtholly lora模型合集

使用数据集:https://huggingface.co/datasets/Carzit/SukaSuka-image-dataset

中珂院炼金学分院QQ群,群号715311859

详细说明:

例如 chtholly-anylora-fanju-20 anylora是训练用的底膜,fanju是训练的数据集使用的多为番剧图片,所以会有很好的番剧效果

统一测试参数为 (大致框架)

建议使用高epoch的lora模型时,减小词相关力度,在4-7之间 (masterpiece, best quality, hi-res:1.2),1girl, solo,upper body,long hair, blue hair, blue eyes, two side up, dress, lora:chtholly-animefull-fanju-20-000010:0.65

ugly,bad anatomy,bad hands,extra hands,missing fingers,broken hand,more than two hands,well proportioned hands,more than two legs,unclear eyes,missing arms,mutilated,extra limbs,extra legs,cloned face,fused fingers,extra digit, fewer digits,extra digits,jpeg artifacts,signature,watermark,username,blurry,large breasts,worst quality,low quality,normal quality,mirror image, Vague

euler a 512 768 30 8

其中lora模型名称更换但是权重为0.65 个人建议出图权重在0.6-0.8之间

因为个人原因,就展示了底膜为anylora时使用这些lora

具体参数可以自己查看,主要使用了epoch为5,10,15,20

lora:chtholly-animefull-fanju-20

lora:chtholly-anylora-fanju-20

因为底膜与训练一致,有一定加成

lora:chtholly-anylora-chatu-20