<h3 id="heading-26">更新</h3>基于Pony模型训练了一版人物lycoris，可适用于pony系模型。<h3 id="heading-27">导言</h3>这是一个《Umamusume Pretty Deby》<s>《闪耀</s><s>优俊少女</s><s>》</s>的里见皇冠角色LyCORIS，目前在AOM3A3上测试良好。建议以0.6左右的权重使用。在使用中使用cut off插件（<a target="_blank" rel="ugc" href="https://github.com/hnmr293/sd-webui-cutoff">https://github.com/hnmr293/sd-webui-cutoff）</a>防止元素污染会获得更好的效果。<h3 id="heading-28">训练细节</h3>以NAI Final作为底模训练，使用了46张图片作为训练集，将约一半图片作为回归训练集，每张图片重复训练10次，训练10个epoch，最终步数为900步，以下为主要设置参数：注意没有镜像增强，因为镜像增强可能对单边的发饰绘画产生影响。network_dim=32network_alpha=32keep_tokens=4conv_dim=4conv_alpha=4lr="1e-4"unet_lr="1e-4"text_encoder_lr="1e-5"batch_size = 2<h3 id="heading-29">Introduction</h3>This is character LyCORIS of Satono Crown from "Umamusume Pretty Deby". Works well on AOM3A3. I recommend use it in a weight of around 0.6. Using the cut off extension (<a target="_blank" rel="ugc" href="https://github.com/hnmr293/sd-webui-cutoff">https://github.com/hnmr293/sd-webui-cutoff</a>) to prevent element contamination in generation can achieve better results.<h3 id="heading-30">Training Details</h3>Using NAI Final as the base model for training, 46 images were used as the training set, and about half of the images were used as the regression training set. Each image was repeatedly trained 10 times, training 10 epochs, and the final number of steps was 900. The following are the main setting parameters:Note that there is no mirror enhancement, as mirror enhancement may have an impact on hair ornament painting.network_ dim=32network_ alpha=32keep_ tokens=4conv_ dim=4conv_ alpha=4lr="1e-4"unet_ lr="1e-4"text_ encoder_ lr="1e-5"batch_ size = 2

Files with SHA256: 41687210d22efb63d347e30149ee4a24cc44298d66a298dd286d75a02062baad