<h3 id="heading-58">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-59">导言</h3>这是一个《Umamusume Pretty Deby》<s>《闪耀</s><s>优俊少女</s><s>》</s>的樱花桂冠角色LyCORIS。在使用中使用cut off插件（<a target="_blank" rel="ugc" href="https://github.com/hnmr293/sd-webui-cutoff）防止元素污染会获得更好的效果。">https://github.com/hnmr293/sd-webui-cutoff）防止元素污染会获得更好的效果。</a>示例图使用了Viewermix模型进行生成。<h3 id="heading-60">使用方法</h3>建议以0.75左右的权重使用。在打标过程中，为隔离元素污染，采用了每个卡面分开设关键词的方法，以下是各个立绘的关键词：私服：sakura_laurel_CA特雷森制服：sakura_laurel_UN胜负服：sakura_laure_SF<s>（打标的时候打漏了一个字母l，但是就这样吧）</s>其他非特殊立绘：sakura_laurel<h3 id="heading-61">训练细节</h3>以NAI Final作为底模训练，使用了约50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1500步。以下为主要设置参数：network_dim=32network_alpha=32keep_tokens=8conv_dim=4conv_alpha=4lr="1e-4"unet_lr="1e-4"text_encoder_lr="1e-5"batch_size = 2<h3 id="heading-62">局限性</h3><ol><li>人物立绘中的不对称设计的还原难度非常大，基本上只能依靠随机抽卡和手动辅助达到一定的还原效果。</li><li>在出示例图的过程中没注意到Additional Network插件失效，一些LyCORIS的层没有输出，可能对最终结果有所影响。</li></ol><h3 id="heading-63">Introduction</h3>This is a LyCORIS of Mejiro Ardan from "Umamusum Pretty Deby". The difference between other similar characters Lora is the addition of Neige Émeraude. Using the cut off plugin（ <a target="_blank" rel="ugc" href="https://github.com/hnmr293/sd-webui-cutoff">https://github.com/hnmr293/sd-webui-cutoff</a> ）to prevent elemental pollution will achieve better results.Example images are generated by Viewermix model.<h3 id="heading-64">Usage</h3>Suggest using a weight of around 0.75.During the marking process, in order to isolate element contamination, a method of setting keywords separately for each card surface was adopted. The following are the keywords for each vertical drawing:Casual clothes：sakura_laurel_CATeresan uniform：sakura_laurel_UNShobofuku：sakura_laure_SF<s>（I missed the latter "l" when tagging）</s>others：sakura_laurel<h3 id="heading-65">Training Details</h3>Using NAI Final as the base model for training, approximately 50 images were used as the training set, and about half of the images were used as the regression training set. Each image was repeated 10 times for 10 epochs, with a final step count of 1500.The following are the main settings parameters:network_ dim=32network_ alpha=32keep_ tokens=8conv_ dim=4conv_ alpha=4lr="1e-4"unet_ lr="1e-4"text_ encoder_ lr="1e-5"batch_ size = 2<h3 id="heading-66">Limitations</h3><ol><li>It is very hard to draw the asymmetrical designs in character portraits. Maybe a lot of tries and manual assistance is a certain way to achieve good results.</li><li>During the process of presenting the example image, I did not notice that the Additional Network plugin was invalid, and some LyCORIS layers did not output, which may have an impact on the final result.</li></ol>

Files with SHA256: 64b1d10dd86f4e66b79cff1ddd935cef88e3adb178ecb4d92f6755330391a290