<h3 id="heading-4133">V3.0 Update</h3>来来回回测试了三个月，总算是有阶段性的成果了改用LyCORIS进行训练，以保留更多的服饰细节，模型需要使用<a rel="ugc" href="https://github.com/KohakuBlueleaf/a1111-sd-webui-lycoris">a1111-sd-webui-lycoris插件</a>加载往数据集里新添加了一些素材，优化打标方式，现在可以将几部分的特征比较好地分离出来，主tag还是<code>nahida \(genshin impact\)</code>，主要负责头部除头饰和瞳孔以外的特征生成，四叶印瞳孔的tag是<code>cross-shaped pupils</code> ，但目前效果不太理想。想使用原版服装的话可以根据需要添加下列词条：<code>in-game dress, hair ornament, bloomers, stirrup legwear, bracelet, cape</code>，想换装就参考图例用其他tag推荐unet权重0.4~1.0，低权重下可能需要手动添加耳朵、发型tag<code>pointy ears, multicolored hair, gradient hair, white hair, green hair</code>来维持人物特征如果感觉画面较脏，推荐使用<a target="_blank" rel="ugc" href="https://civarchive.com/models/22354/clearvae">ClearVAE</a>改善画质Translate with DeepL:After three months of testing back and forth, we finally have a milestone result. Change to use LyCORIS for training to retain more details of clothing, this version of model needs to be loaded using <a rel="ugc" href="https://github.com/KohakuBlueleaf/a1111-sd-webui-lycoris">a1111-sd-webui-lycoris plugin</a>. To the data set to add some new material, optimize the tagging method, now can be better separated from the features of several parts, the main tag is still <code>nahida \(genshin impact\)</code>, mainly responsible for the head in addition to headdress and pupils other than the generation of features, four-leaf print pupils tag is <code>cross-shaped pupils</code>, but the current effect is not ideal. If you want to use the original clothing, you can add the following tags as needed: <code>in-game dress, hair ornament, bloomers, stirrup legwear, bracelet, cape</code>, want to change the reference legend with other tag.Recommended unet weight 0.4~1.0, low weight may need to manually add ears, hairstyle tags <code>pointy ears, multicolored hair, gradient hair, white hair, green hair</code> to maintain character features.If you feel that the picture is a little dirty, we recommend using <a target="_blank" rel="ugc" href="https://civarchive.com/models/22354/clearvae">ClearVAE</a> to improve the quality. <h3 id="heading-4134">V2.0 Update</h3>I learned a lot in this month, and my graphic card also suffered a lot :P, till now I can publish a better version for this modelFor this version, recommend weight is 0.45~0.7, only one trigger word <code>nahida \(genshin impact\)</code> is needed, but has better restoration of characteristicMy suggestion is use lower weight for model, and higher weight for trigger word, so you can easily change her suitAll known problems in previous version are solvedThe next version will be a LyCORIS model, please stay tuned!Any suggestions are welcome!<h2 id="heading-4135">V1.0 Release</h2>Genshin Impact Nahida LoRA, training with base model final-prune[925997e9].Recommend weight 0.45~0.8. around 0.7 is a sweet spot for customize and keep characteristic.Models used for examples:<ul><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/6755/cetus-mix">Cetus-Mix</a></li><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/7240/meinamix">MeinaMix</a></li><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/15509/aoaoko-pvc-style-model">AOAOKO [PVC Style Model]</a> with VAE <a target="_blank" rel="ugc" href="https://civarchive.com/models/4468/counterfeit-v25">Counterfeit-V2.5</a></li></ul>Use tags <code>braid, sidelocks, side ponytail, multicolored hair, white hair, green hair, gradient hair</code> to fix hair style.Use tag <code>bracelet</code> to generate a in-game style bracelet (working better in higher weight).Use tag <code>white bloomer</code> to add in-game style bloomer for charater.Known problem:<ul><li>Lower weight may cause a little incorrect on the green part of the hair color.</li><li>Sometimes a unexpected crown would appear on Nahida's head (mostly in higher weight).</li></ul>Any suggestions are welcome!

Nahida Genshin Impact

this version is trained by LyCORIS network, you need to download <a rel="ugc" href="https://github.com/KohakuBlueleaf/a1111-sd-webui-lycoris">plugin a1111-sd-webui-lycoris</a> to load it

Files with SHA256: 6565f718d90d291033d22d4e2e3f84a4c0ac54b1dfa912c53476098f40d4636a