<h2 id="model-description">Model Description</h2>This LoRA model is a character enhancement model for The Dahlia, fine-tuned on a dataset of 500+ images. Compared to the previous version, the dataset images have been cut to reduce overfitting to lighting and shadows.In addition to the version based on Illustrious XL, I have also trained a version based on ZukiAnimeIll (a derivative model of Illustrious XL). Users employing ZukiAnimeIll for inference can use either this version or that specific one.<h2 id="usage-tips">Usage Tips</h2><ul><li>Clothing &amp; Accessories: Special optimization has been applied to address the hard-baked clothing issue. The Dahlia's official accessories are now detachable, bound to the tags: <code>"hat, dress, gloves, high heels, tail"</code>.<ul><li>Omitting these tags will default to her official illustration appearance.</li><li>Note: The <code>"gloves"</code> tag defaults to one black and one white glove.</li></ul></li><li>Changing Outfits: During testing with my usual checkpoint models, removing the official clothing tags allowed for adding any preferred clothing elements. If residual elements persist when using your specific checkpoint for inference, consider adding the aforementioned clothing tags to the negative prompt.</li><li>Anatomy Balance: Due to the character's naturally large breast size combined with the influence of certain checkpoint models, the results may sometimes appear overly bulky. To balance this, consider adding <code>"large breasts"</code> to the negative prompt or <code>"medium breasts"</code> to the positive prompt.</li></ul><h2 id="recommended-settings">Recommended Settings</h2><ul><li>Recommended weight: 0.7 - 1.0</li><li>(Please adjust the prompts according to the official documentation of the checkpoint model you are using.)</li><li>upper-body prompt :</li></ul><pre><code>masterpiece, best quality, newest, absurdres, highres, 1girl, solo, Constance, hat, dress, detailed light, epic, dramatic, flaming eye, close view, close up, finger to mouth,
</code></pre><ul><li>full-body prompt :</li></ul><pre><code>masterpiece, best quality, newest, absurdres, highres, 1girl, solo, Constance, hat, dress, gloves, high heels, tail, standing, medium breasts, detailed light, epic, dramatic, 
</code></pre><h2 id="others">Others</h2>While I am very satisfied with the accuracy of The Dahlia's features, including her distinctive corset and gloves (features not consistently achieved in other LoRAs I've tested), I plan to further optimize the details after the official 3D model is released, provided I have the spare time.If you have any feedback, please feel free to leave comments so I can make adjustments. Thank you.Bilibili账号：<a target="_blank" rel="ugc" href="https://space.bilibili.com/12595237?spm_id_from=333.1007.0.0">https://space.bilibili.com/12595237?spm_id_from=333.1007.0.0</a>小红书账号：<a target="_blank" rel="ugc" href="https://www.xiaohongshu.com/user/profile/670412f2000000001d0209cd">https://www.xiaohongshu.com/user/profile/670412f2000000001d0209cd</a><h3>模型简介</h3>这个LoRA模型是角色 大丽花 (The Dahlia) 的角色增强模型，在 500+ 张图片的数据集上微调得到。 与前一版相比，主要区别在于对数据集进行了瘦身，减少了对光影的过拟合。 除了基于Illustrious XL的版本，我还训练了基于ZukiAnimeIll（Illustrious XL的一款衍生模型）的版本，使用ZukiAnimeIll推理的用户可以使用该版，也可以使用那一版。<h3>使用建议</h3><ul><li>服装与配饰 ： 针对服装 hard-baked问题进行了专门优化。现在大丽花的官方外设是可拆卸的，绑定了以下标签：<code>hat, dress, gloves, high heels, tail</code><ul><li>如果不添加这些服装提示词，模型将默认呈现官方立绘造型。</li><li>注： <code>gloves</code> 标签默认包含一黑一白的手套特征。</li></ul></li><li>更换服装： 实测下来我在我常用的checkpoint模型上在去掉上述官方服装提示词后，可以自由添加任意喜欢的服装元素。如果你在使用其他checkpoint模型推理时仍发现有官方服装残留，建议将上述标签（如 <code>dress</code>, <code>hat</code> 等）加入到 负面提示词 (Negative Prompt) 中。</li><li>人体结构平衡 ： 由于角色设定原本胸部尺寸较大，加上某些Checkpoint模型的影响，可能会导致胸部尺寸过于臃肿。如果遇到该情况可尝试在 负面提示词 中加入 <code>large breasts</code>；或者在 正面提示词 中加入 <code>medium breasts</code> 以进行平衡。</li></ul><h3>推荐设置</h3><ul><li>推荐权重： 0.7 - 1.0</li><li>（请根据你实际使用的Checkpoint模型的官方文档，适当调整提示词策略。）</li><li>上本身prompt:</li></ul><pre><code>masterpiece, best quality, newest, absurdres, highres, 1girl, solo, Constance, hat, dress, detailed light, epic, dramatic, flaming eye, close view, close up, finger to mouth,
</code></pre><ul><li>全身prompt:</li></ul><pre><code>masterpiece, best quality, newest, absurdres, highres, 1girl, solo, Constance, hat, dress, gloves, high heels, tail, standing, medium breasts, detailed light, epic, dramatic, 
</code></pre><h3>小记</h3>虽然我已经对大丽花的特征还原程度很满意，包括她独特的胸衣，手套等（目前我试用过其他的LoRA没有在这些特征做到一致），但我如果有课余时间，会在官方3d模型发布后针对细节优化。如果有任何反馈意见欢迎评论，以便我做出调整，谢谢。

Files with SHA256: d568337fcf63add77687aed837a4c375c2b0dcd00b8a25da9c2f153e3a3c35bf