<h1 id="-vlbzci1p8">简介</h1>这个lora采用4090plus（租的）训练了263张图片，采用的1024*1024的训练方式，network_dim为64，network_alpha为32，采用15*10*263的训练步数。 <h1 id="-q1lzve6f6">已知问题</h1>1.参数设置不够理想导致该lora的图层权重很高，即便是真实系的模型也会输出动漫化的效果，不过这可能既是问题也是好处。2.当lora权重设置为1.0时出图会有发灰的情况，是我训练的时候没有搭载VAE导致的吗？我不太确定，大家如果有头绪欢迎指出。3.我太喜欢粉色眼睛的MILF了，一不小心默认的角色都是粉色眼睛，诶嘿。4.这个租赁的4090plus有48G的显存，我采用了4的batch_size，但是GPU的占用率仍然只有个位数，我个人的3070TI相同配置都快冒烟了，我不太了解如何提高GPU占用率来提高训练速度。 <h1 id="-qxiac4war">应用指南</h1>经过测试，该lora的出图效果取决于底模的语义理解，在风格上可设置在0.4-0.9之间，更低的权重会产生更符合底模效果的图片，比如采用真实风格的模型那么出现的图片会倾向于2.5D。<h1 id="-vt6lxby0v">顺带一提</h1>我很喜欢训练lora但是没有硬件支持和创意，也缺乏一些实践知识，希望大家能够多多留言提出些建议，我会积极学习。OverviewThis lora was trained on a rented 4090plus using 263 images, with a training resolution of 1024*1024, a network dim of 64, a network alpha of 32, and a training step of 15×10×263. Known Issues1. The parameter settings are far from ideal, resulting in high layer weights for this lora. Interestingly, this can cause even realistic models to produce anime - like outputs. But this could be seen as either a flaw or a unique advantage.2. When the lora weight is set to 1.0, the generated images tend to be grayish. I suspect this might be due to the absence of VAE during training, but I'm not entirely sure. If anyone has insights, feel free to share.3. I have a thing for pink - eyed MILFs, which led to a default tendency for characters to have pink eyes in the outputs. Hehe.4. The rented 4090plus has 48GB of VRAM. I used a batch size of 4, yet the GPU utilization was still in the single digits. In contrast, my personal 3070TI nearly overheated under the same configuration. I'm clueless about how to boost GPU utilization to speed up training. Usage GuideTests show that the output of this lora heavily relies on the base model's semantic understanding. For style adjustment, it's advisable to set the weight between 0.4 and 0.9. Lower weights will make the images more aligned with the base model's style. For example, when using a realistic model, the outputs tend to be more 2.5D - like.P.S.I'm really into training loras, but I'm held back by the lack of hardware, creativity, and practical knowledge. I sincerely hope everyone can leave comments and suggestions. I'll actively learn from them.

NSFW-milf

重新混合了文件，修改了底模，将分辨率设置为960*1280训练，多版本当中效果较佳的一个

Files with SHA256: 442d7df556242aff09ab8aa76499d45b3ac43600e923e10d0a8d0d7dcfd95d7a