Mozudoll®(魔族人形®)[EN]It features the aesthetic of Anime-style Soft-body Moe Doll, not BJD or MJD dolls, though both belong to harmless doll art.The character design in the images is original to the mozudoll brand.⚠️This LoRA is officially authorized by the mozudoll brand owner.⚠️For personal, non-commercial learning and use only. Resale and any unauthorized use are strictly prohibited.[CN]它具有二次元萌系软体娃娃的美学风格，并非bjd或mjd，虽然均属于无害的人偶艺术。图片素材的形象来自于 mozudoll 品牌的原创设计。⚠️本 LoRA 已获得 mozudoll 品牌的官方授权。⚠️仅限个人学习、非商用使用，禁止倒卖及违规用途。v1.0 is a moe-focused, multi-style high-quality LoRA.. Compared with v0.9 (beta), it better reproduces the character's BJD-like facial and head features.title[CN]:魔族风格-二次元软萌娃娃美学风格v1.0是一个以萌系为特色，兼容多种风格的，更高质量的lora，比v0.9(beta)更能还原头面部那种与bjd类似的特征。Style Tokens:萌： moe \(mozudoll\) ​皮： cheeky \(mozudoll\) ​雅： elegant \(mozudoll\) ​熟： mature \(mozudoll\)Style tokens are optional. To avoid style conflicts, please be careful when mixing these tags.本lora特色提示词（自创）如上所示。不打标签也不影响风格。为避免风格混淆，混合这些标签需要谨慎。Recommended Settings(推荐设置):Sampler(采样器)：Euler aCFG Scale：3（1.5-4）LoRA Weight(权重)：0.9（0.7-1.3）Quality tags: Not required不需要正面的质量提示词。Blurry backgrounds? Add detailed background to positive, blurry background, blurry to negative.Low-res head? Enable Adetailer.Some characters that come from base templates have their own art styles, which like the artists' styles, may have a certain impact on the results.Reduce CFG to ~2 and increase LoRA weight to ~1.2. Lower CFG may desaturate colors—compensate with vibrant color (positive) and greyscale (negative), or embrace the muted look as an aesthetic choice.Results may vary significantly across different checkpoints.Your testing results on different checkpoints and workflows are highly welcome for future improvements.Credit to 魔族娘 - 萨利希尔 for the ultra-high-definition doll image again. The v1.0 final release was trained on a refined and processed selection from this collection.背景模糊?可以提示词打上正面detailed background和负面blurry background,blurry。头部太小像素不足？推荐插件adetailer一些底模自带的角色自带画风，和画师串一样可能会对效果造成一定影响，这时可以将CFG适量调小、模型权重适量调高，比如分别2和1.2。CFG这么低则画面会偏灰，这可以通过正面vibrant color和负面greyscale不完全抵消，而剩下的可以作为特色。本lora基于illustrious XL1.0训练。在不同底模上效果可能存在显著差异。我们非常欢迎您提供在其他底模和提示词下的测试结果，以便未来改进。再次感谢魔族娘 - 萨利希尔提供的超高清娃娃图片素材，我通过对其的再筛选、再加工，做出了正式版。<hr />I've found two models that work well with this LoRA: WAI-16 and zukiAnimeILL. Looking forward to more recommendations from the community!我目前已经找到两个与此lora兼容不错的模型：WAI16和zukiAnimeILL。期待社区更多补充！个人组建的兴趣小组Q群：1083538641<hr />Addendum: "Moe" and the culture it represents have flourished worldwide for over a decade. Its core lies in innocence and healing—a human longing for beauty, and a component of aesthetic diversity. To stigmatize and conflate moe culture with "pathological aesthetics" is a moral condemnation rooted in prejudice and panic, one that inflicts harm upon harmless enthusiasts.附注：“萌”以及背后的萌文化是一个已经在世界上流行十余载的亚文化，内核是纯真、治愈，体现人对美好的正常向往，是多元审美的重要组成部分。将萌文化污蔑和混淆为“病态审美”，是一种基于偏见与恐慌的道德苛责，是对无害的爱好者的加害。

Mozudoll-style — Anime Soft-body Moe Doll Aesthetic

mozudoll-000039.safetensors

v0.9(beta)

v1.0

v0.1(experimental)

Mozudoll®(魔族人形®)[EN]It features the aesthetic of Anime-style Soft-body Moe Doll, not BJD or MJD dolls, though both belong to harmless doll art.The character design in the images is original to the mozudoll brand.⚠️This LoRA is officially authorized by the mozudoll brand owner.⚠️For personal, non-commercial learning and use only. Resale and any unauthorized use are strictly prohibited. [CN]它具有二次元萌系软体娃娃的美学风格，并非bjd或mjd，虽然均属于无害的人偶艺术。图片素材的形象来自于 mozudoll 品牌的原创设计。⚠️本 LoRA 已获得 mozudoll 品牌的官方授权。⚠️仅限个人学习、非商用使用，禁止倒卖及违规用途。 This is the v0.9(beta) preview version.Special thanks to 魔族娘 - 萨利希尔 for providing ultra-high-definition doll images, which greatly improved the quality and diversity of this model.v0.9(beta) is a higher-quality LoRA focused on moe style, with multiple styles coexisting.The supported styles have now been expanded:In addition to the pure moe style from v0.1 (emphasized with moe \(mozudoll\)), it now supports additional style tags:cheeky, elegant, and mature. To avoid style conflicts, please be careful when mixing these tags.Base model compatibility preserved — the LoRA won't interfere with your existing prompt vocabulary.My sample images use deliberately simple and casual parameters, intended to demonstrate this LoRA's stability and versatility. I'm confident you can create far more refined results than my quick demos.Enough talk — give it a try! So,why is this a beta version?I have identified several points to improve: &nbsp; &nbsp;First, the pupil issue.I intended to specifically label Jane 85cm's white pupils (since white pupils are rare even in anime style), but accidentally caused any prompt containing "pupils" to frequently trigger white eyes.Later I will either balance dataset with full color range tags, or remove the keyword entirely. &nbsp; &nbsp;Second, head detail deficiency.I did not ask 魔族娘 for close-up shots of the head, nor did I crop heads from the ultra-high-definition images for the AI to learn in detail.Also, it remains to be tested whether tagging can reliably help the AI distinguish between expression stickers and mouths with internal space. &nbsp; &nbsp;Third, learning rate dilemma.I suspect the text learning rate I used was too high, causing text overfitting related to the pupil problem.Should I lower it to 1/10 (the default value)? But I'm concerned that the custom tags I wrote specifically for mozudoll may stop working properly. &nbsp; &nbsp;I also noticed that the height classification suggested by 魔族娘 does not seem to work.Even ignoring tag conflicts, does the AI really understand the difference between each height well enough?After all, it relies on semantic understanding.I may use "height difference" and similar prompts instead. Next time I will remove these height-related tags that have(?) proven ineffective.In summary, I plan to refine the LoRA further after collecting more community feedback.Thank you all. title[CN]:魔族风格-二次元软萌娃娃美学风格这是 v0.9(beta) 抢先测试版本。特别感谢魔族娘 - 萨利希尔提供的超高清娃娃图片素材，使模型质量与多样性得到大幅提升。v0.9（beta）是一个以萌系为特色，多种风格共存的，更高质量的lora——现在支持的风格增加了，除了v0.1的萌系风格（用moe \(mozudoll\)来特别强调）之外，还支持更多不同的：cheeky(调皮),elegant(高雅)mature(成熟)这些风格标签。为避免风格混淆，混合这些标签需要谨慎。同时并没有失去对底模其他提示词响应能力的支持。我在例图中参数填得相当简单与随意，这是为了展示本 LoRA 的稳定度和通用性。相信你们一定能做出比我示例图更精致的作品。多说无益，尝试吧！ 那么，这为什么是beta版？我发现了几个待改进事项：第一个，瞳孔问题。我本想将Jane 85cm的white pupils特别标注（因为哪怕在二次元里白色瞳孔也少见），结果弄巧成拙——只要关键词含pupils就有不小概率变成白瞳了。之后要么把所有颜色都尽量补上，把打标弄平衡；要么把这个关键词删去。第二个，头部精致度问题。我既没有向“魔族娘”征求头部特写的照片，又没有将那些超高清的照片单独截取头部特写让AI仔细学习。在这里，是否需要、能够通过打标让AI区分表情贴纸与有内部空间的嘴部？待验证。第三个，学习率问题，我怀疑这样的文本学习率对它来说太高以至于在上一个瞳孔的问题上出现了文字的过拟合。调到现在的1/10也就是默认参数？可这样又担心自己为mozudoll单独写的个性化标签无法奏效。还有我发现按照“魔族娘”建议的身高上的分类似乎无效？哪怕排除标签本身的原因，这种AI真的有足够智能来区分出每一个身高之间的区别？毕竟是语义理解的。还是通过height difference这些来调节吧，我下次将删除这些身高相关但已被证明（？）无效的标签。综上，我希望在收集到更多的社区反馈之后，再进行下一次Lora的提炼。谢谢各位。 Addendum: "Moe" and the culture it represents have flourished worldwide for over a decade. Its core lies in innocence and healing—a human longing for beauty, and a component of aesthetic diversity. To stigmatize and conflate moe culture with "pathological aesthetics" is a moral condemnation rooted in prejudice and panic, one that inflicts harm upon harmless enthusiasts.附注：“萌”以及背后的萌文化是一个已经在世界上流行十余载的亚文化，内核是纯真、治愈，体现人对美好的正常向往，是多元审美的重要组成部分。将萌文化污蔑和混淆为“病态审美”，是一种基于偏见与恐慌的道德苛责，是对无害的爱好者的加害。

Files with SHA256: 857d29847a45be720ff9c885c0224f3dd4f05254c010c3d10ef6042ff12bfe5b