<p>使用约800张图片进行训练，图片内容为 70%官方图片，20%AI高质量图片，10%同人图片。使用animagineXLV3_v30进行训练，002以后的有些过拟合。训练分辨率为1024*1024（在a40显卡上训练了14个小时 (还失败了...悲...))。触发提示词为nekopara__azuki,其余触发词和我发布的1.5的azuki lora是一样的，（sdxl的训练数据集就是在1.5版本的数据集上做了超分增强等操作...增加了图片...<br /> training was conducted using approximately 800 images, consisting of 70% official images, 20% high-quality AI images, and 10% fan-made images. The training utilized the animagineXLV3_v30, noting some overfitting issues with versions after 002. The training resolution was set at 1024*1024, and it was conducted on an A40 graphics card for 14 hours, though it ended in failure (sadly...). The trigger keyword used was nekopara__azuki, and the rest of the trigger words are the same as those released in my 1.5 version of azuki lora. The training dataset for sdxl is essentially an enhanced version of the 1.5 dataset, with super-resolution and other techniques used to add more images.</p>

Files with SHA256: 374d5aa5755fee80746d2bf58d01605f3c7a83844a83bad8b4c87f9677078696