[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。Recommended Settings(推荐设置):Sampler(采样器)：Euler aQuality tags: Not required不需要正面的质量提示词。Special NoticeWhen using this model, recognition of certain characters may be impaired. To resolve this issue, use escape characters for emphasis.Example: momoi (blue archive) → momoi \(blue archive\)Also,lowering the LoRA weight from 1.0 to 0.6–0.8 can further improve the feature response of these characters.Additional NotesThis is an experimental release.It took 38 hours for me to make this LoRA purely on a trial run, using WD14 for automatic tagging. To my surprise, the training dataset was of such high quality that the model turned out pretty well, so I decided to share it as an experimental build.I unilaterally removed the backslash escape character \ during preprocessing. This led to a degradation in style classification performance (I will definitely correct it next time), but the native styles of characters recognized by the base model still function normally.I also added extra prefix tags: realistic, anime doll. In hindsight, these tags appear to be redundant? The model works perfectly fine without them during image generation, and they likely became distractions during training.I know that tagging with large language models (LLMs) would produce better results, but I'll leave this work for the next version.Lastly, I accidentally typed anime instead of anima out of habit when naming the model — please pardon this silly mistake.特别提醒：使用该模型时，对部分角色的响应会变差，此时可以通过转义符来强调，如：momoi (blue archive)→momoi \(blue archive\)还有将lora权重从1降到0.6～0.8，可以进一步改善这些角色的特征响应。附注：这是一个实验性版本。我抱着试一试的态度，拿着wd14的打标花了38个小时炼制了这个lora。没想到由于素材本身质量够高，效果不错，于是当作实验版本发布。我当时擅自删去了转译用的反斜杠"\"，这会导致风格的分类效果稍差（我下次一定改正），但底模认识的角色自带的风格仍然是有效的。我当时还擅自加了前缀（realistic, anime doll）但这些看上去似乎是多余的？实际生图时不加这几个标签仍然能正常生成，而这标签反倒很可能成了训练的干扰源。我知道可以拿大语言模型进行更好的打标，但这个工作还是交给下一版本吧。我还在为模型取名时出于习惯把anima打成anime了，请莫见笑。————————个人组建的兴趣小组Q群：1083538641

Mozudoll-Anima

v1.0(experimental)[EN]When using this model, recognition of certain characters may be impaired. To resolve this issue, use escape characters for emphasis.Example: momoi (blue archive) → momoi \(blue archive\)Also,lowering the LoRA weight from 1.0 to 0.6–0.8 can further improve the feature response of these characters.Additional NotesThis is an experimental release.It took 38 hours for me to make this LoRA purely on a trial run, using WD14 for automatic tagging. To my surprise, the training dataset was of such high quality that the model turned out pretty well, so I decided to share it as an experimental build.I unilaterally removed the backslash escape character \ during preprocessing. This led to a degradation in style classification performance (I will definitely correct it next time), but the native styles of characters recognized by the base model still function normally.I also added extra prefix tags: realistic, anime doll. In hindsight, these tags appear to be redundant? The model works perfectly fine without them during image generation, and they likely became distractions during training.I know that tagging with large language models (LLMs) would produce better results, but I'll leave this work for the next version.Lastly, I accidentally typed anime instead of anima out of habit when naming the model — please pardon this silly mistake.[CN]使用该模型时，对部分角色的响应会变差，此时可以通过转义符来强调，如：momoi (blue archive)→momoi \(blue archive\)还有将lora权重从1降到0.6～0.8，可以进一步改善这些角色的特征响应。附注：这是一个实验性版本。我抱着试一试的态度，拿着wd14的打标花了38个小时炼制了这个lora。没想到由于素材本身质量够高，效果不错，于是当作实验版本发布。我当时擅自删去了转译用的反斜杠"\"，这会导致风格的分类效果稍差（我下次一定改正），但底模认识的角色自带的风格仍然是有效的。我当时还擅自加了前缀（realistic, anime doll）但这些看上去似乎是多余的？实际生图时不加这几个标签仍然能正常生成，而这标签反倒很可能成了训练的干扰源。我知道可以拿大语言模型进行更好的打标，但这个工作还是交给下一版本吧。我还在为模型取名时出于习惯把anima打成anime了，请莫见笑。

Files with SHA256: dadce0f4dafa503b06ea97a16c69db13b2aba2dc3fdd7795c17f893b8520fd0f