✨ 极致的京阿尼美学，献给所有热爱光影的创作者。 本 LoRA 旨在完美复现《紫罗兰永恒花园》中令人窒息的顶级动画摄影美学。无论你是想为心爱的二次元角色披上绝美的光影外衣，还是想生成极具电影感的宏大场景，这个模型都能为你提供降维打击般的视觉体验。 ✨ The ultimate Kyoto Animation aesthetics, dedicated to all creators who love light and shadow. This LoRA aims to perfectly replicate the breathtaking, top-tier cinematography and aesthetics of Violet Evergarden. Whether you want to dress your favorite anime characters in gorgeous lighting or generate cinematic, grand landscapes, this model offers a groundbreaking visual experience.🔑 核心触发词 / Core Trigger Word: <code>violet cinematography</code><h3 id="v2-about-v2-version">🆕 关于 V2 版本 / About V2 Version</h3>💡 开发者寄语：V2 并不是传统意义上的简单升级，而是一次全新的技术尝试。因为是初次进行这种混合打标实验，所以最终效果可能没有想象中那么完美无缺，只能说日后会继续多多探索与优化！💡 Developer's Note: V2 is not a traditional upgrade in the conventional sense, but a brand-new technical experiment. Because this hybrid tagging is a first-time experiment, the final output might not perfectly match my highest expectations yet. I will continue to explore and refine this methodology in the future!核心更新： 本次 V2 版本成功将光影效果和材质效果进行了剥离，使得它们不再与触发词强关联，避免了所有画面都被强制渲染出过于浓郁的“紫罗兰味”。但是！在增加画面可控性的同时，这也相应地提高了提示词（Prompt）的编写难度。Core Updates: In this V2 version, I have successfully decoupled the lighting, shadows, and material effects from the trigger word. This prevents the "Violet" style from forcibly overpowering every generation. The result? Greatly enhanced controllability, though it comes with a trade-off: it raises the difficulty of prompt engineering.训练集扩充： 本次大幅扩大了训练集，完整收录了十二集番剧和两部剧场版，训练集总计高达 6234 张剧照。若是以后修改训练方向或者进一步优化训练集，恐怕 6234 这个数量就已经是上限了。Dataset Expansion: The dataset has been massively expanded to include all 12 TV episodes and 2 theatrical movies, totaling 6,234 high-quality screencaps. Even if I adjust the training direction or further optimize the dataset in the future, this 6,234 figure is likely the absolute ceiling for the dataset size.全新打标逻辑： 本次初次采用了 WD14 单词打标 + 千问自然语言混合打标的方案： Pioneering Tagging System: For the first time, this model utilizes a hybrid tagging method: WD14 Tags + Qwen Natural Language:<ol><li>WD14 单词标的作用是“挡枪”： 把我们不需要模型学的东西（特定的衣服、动作、长相）用单词解释掉，告诉模型“这部分你底模自己懂，别把它们算进我的新画风里”。</li><li>WD14 Tags (The "Shield"): These act as a filter. By using tags to describe specific clothes, poses, and facial features, we tell the base model, "You already know these elements, don't bake them into my new style."</li><li>千问自然语言的作用是“点睛”： 用高阶的摄影术语，把底模平时很少见到的“极品光学现象”框选出来，告诉她“好好看，好好学”。</li><li>Qwen Natural Language (The "Highlighter"): This uses advanced photography terminology to frame the extreme optical phenomena rarely seen by the base model, guiding the AI to focus on and learn these specific rendering techniques.</li></ol>分类类型与打标侧重： 本次训练集被精细划分为四大类，分别对应不同的训练目的： Dataset Categorization &amp; Training Focus: The dataset was divided into four specialized categories, each with a unique tagging focus:<ul><li>A. 单人图 (人脸+全/半身)： 主要目的：锚定角色长相、服装细节。打标侧重：WD 负责角色触发词+物理细节；千问负责光影渲染+皮肤材质。</li><li>A. Single Person (Face/Full/Half body): Anchors character appearance and clothing details. WD: Character trigger words + physical details. Qwen: Lighting rendering + skin materials (subsurface scattering).</li><li>B. 多人交互图： 主要目的：锚定构图、空间关系、站位。打标侧重：WD 负责交互动作、机位标签；千问负责景深、透视、三维空间感。</li><li>B. Multiple People (Interaction): Anchors composition, spatial relations, and positioning. WD: Interaction actions + camera angles. Qwen: Depth of field, perspective, and 3D spatial awareness.</li><li>C. 大场景 (无人)： 主要目的：锚定画风、空间氛围。打标侧重：WD 负责地理概念词；千问负责宏观构图、气候气象、空间纵深。</li><li>C. Scenery (No humans): Anchors the overall style and atmospheric environment. WD: Geographical concept words. Qwen: Macro composition, weather/climate, and spatial depth.</li><li>D. 特写 (物品/手脚)： 主要目的：锚定微观材质、复杂结构。打标侧重：WD 负责物品名称；千问负责微观纹理、金属/布料质感、焦距表现。</li><li>D. Close-up (Objects/Hands/Feet): Anchors micro-materials and complex structures. WD: Object names. Qwen: Micro-textures, metal/fabric textures, and focal expression.</li></ul>核心光影词汇集： 整个打标方案的核心光影词汇集约为 15 个高频标准术语，配合 4 个文件夹各自的特化方向，构成了一套面向 LoRA 画风训练的技术性自然语言标签体系：Core Vocabulary: The core lighting and shadow vocabulary for the entire tagging scheme consists of about 15 high-frequency standard terms, tailored to the 4 folders to form a technical natural language tag system for LoRA style training:<ul><li>光照类 (Lighting): volumetric lighting, soft bloom, light diffusion, specular highlights</li><li>材质类 (Materials): translucent shading, subsurface scattering, material response, surface texture</li><li>空间类 (Spatial): depth layering, atmospheric perspective, feathered edges</li><li>绘画技法类 (Techniques): painterly rendering, layered color washes, soft gradient blending</li></ul><h3 id="(v1-)-hardcore-training-process-(v1-foundation)">🛠️ 硬核的训练过程 (V1 版本纯正打底) / Hardcore Training Process (V1 Foundation)</h3>💡 注：以下为 V1 版本的初期硬核提炼说明。正是基于 V1 如此纯粹的画风打底，才有了如今 V2 版本的庞大扩充：💡 Note: The following describes the initial hardcore extraction process for the V1 version. It is upon this extremely pure stylistic foundation that the massive V2 expansion was built:为了保证画风的绝对纯正，我完全摒弃了网络上的杂乱同人图，采用了“纯正原剧提取”的硬核炼丹方式： To ensure absolute stylistic purity, I completely discarded messy fan art from the internet and adopted a rigorous "pure original anime extraction" method:<ul><li>百里挑一： 以每两秒一次的频率从动画原剧中截图，在近 10,000 张原始素材中，人工精选出 1,000+ 张构图最完美、光影最极致的画面。</li><li>Carefully Handpicked: Extracted frames every two seconds from the original anime, manually selecting the best 1,000+ images with perfect composition and extreme lighting from nearly 10,000 raw screenshots.</li><li>全要素覆盖： 训练集不仅包含人物特写，还涵盖了精美静物与宏观场景（人、物、景），确保模型理解“整个世界”。</li><li>All-Encompassing: The dataset includes character close-ups, exquisite objects, and grand sceneries (characters, props, backgrounds) so the model understands the "whole world."</li><li>细致打标： 对光影类型、镜头角度、材质反光进行了极为细致的人工打标工作。</li><li>Meticulous Tagging: Highly detailed tagging for lighting types, camera angles, and material reflections.</li></ul><h3 id="model-highlights">🌟 模型亮点 / Model Highlights</h3><ul><li>极致光影 (Cinematic Lighting)： 完美继承了丁达尔体积光、水面焦散折射、环境漫反射等高级光影表现。</li><li>Cinematic Lighting: Perfectly inherits advanced lighting effects like Tyndall volumetric light, water caustics, and ambient diffuse reflection.</li><li>绝佳横图 (Excellent Landscape Shots)： 针对横图与宽画幅构图进行了深度优化，生成大场景时极具电影截帧质感。</li><li>Excellent Landscape Shots: Deeply optimized for landscape and wide-format compositions, generating cinematic screencap textures for grand sceneries.</li><li>超强泛化 (High Generalization)： 无论是原著角色还是套用其他顶级流量 IP 角色，都能完美融合，画风表现极其稳定。</li><li>High Generalization: Whether using original characters or applying other top-tier IP characters, it blends perfectly with extremely stable style performance.</li></ul><h3 id="usage-tips">💡 推荐参数与使用建议 / Usage Tips</h3>推荐权重 / Recommended Weight: ~ 0.8提示词建议 / Prompting Advice: 建议使用自然语言长句描述画面，避免单纯的碎标签堆砌。模型可玩性极佳，范例提示词可以参考示例图片，或者自行探索！ Highly recommend using natural language sentences instead of comma-separated tags. The model has excellent playability; please refer to the example images for prompt ideas or freely explore on your own!<h3 id="example-prompt">📝 范例提示词 / Example Prompt</h3><code>violet cinematography, A gorgeous bright beach scene of Violet Evergarden walking barefoot in the shallow water of a crystal-clear ocean. She is wearing her classic blue and white postal uniform along with her signature brown leather gloves, with her blonde braided hairbuns and glowing emerald brooch standing out beautifully. The bright midday sun creates shimmering, blinding specular reflections on the rippling water surface and casts beautiful light caustics that highlight the wet skin on her legs</code><h3 id="share-your-art!">📸 求返图！交流学习 / Share Your Art!</h3>最后，跪求各位大佬多多返图！非常期待看到大家用这个模型玩出新花样，更希望能借此机会，学习一下各位大佬们出神入化的提示词功底！ Finally, please share your generated images in the reviews! I am really looking forward to seeing how you play with this model, and I would absolutely love to learn from your god-tier prompt engineering skills!<h3 id="disclaimer">⚠️ 免责声明 / Disclaimer</h3>本模型仅供 AI 绘画爱好者学习、交流与同好分享使用，严禁用于任何形式的商业用途。本模型旨在向京都动画（Kyoto Animation）极致唯美的美术风格致敬，无任何恶意侵权意图。《紫罗兰永恒花园》的所有相关角色设计、美术风格及知识产权均归 京都动画 (Kyoto Animation) 及 晓佳奈 (Kana Akatsuki) 所有。使用者由此产生的一切商业纠纷或法律责任由使用者自行承担，模型作者概不负责。 This model is strictly for educational purposes, personal communication, and sharing among fans. Any form of commercial use is strictly prohibited. This model is created as a tribute to the exquisitely beautiful art style of Kyoto Animation, with absolutely no malicious intent to infringe upon copyrights. All character designs, art styles, and related intellectual properties of Violet Evergarden belong to Kyoto Animation and Kana Akatsuki. The creator of this model assumes no responsibility for any commercial disputes or legal issues arising from its use.

Violet Evergarden Artstyle (紫罗兰永恒花园风格) | Kyoto Animation (京都动画)

Violet LoRA V2 Release NotesTrigger Word: <code>violet cinematography</code>Core Updates: In this V2 version, I have successfully decoupled the lighting, shadows, and material effects from the trigger word. This prevents the "Violet" style from forcibly overpowering every generation. The result? Greatly enhanced controllability, though it comes with a trade-off: it raises the difficulty of prompt engineering.Dataset Expansion: The dataset has been massively expanded to include all 12 TV episodes and 2 theatrical movies, totaling 6,234 high-quality screencaps. Even if I adjust the training direction or further optimize the dataset in the future, this 6,234 figure is likely the absolute ceiling for the dataset size.Pioneering Tagging System: For the first time, this model utilizes a hybrid tagging method: WD14 Tags + Qwen Natural Language.<ol><li>WD14 Tags (The "Shield"): These act as a filter. By using tags to describe specific clothes, poses, and facial features, we tell the base model, "You already know these elements, don't bake them into my new style."</li><li>Qwen Natural Language (The "Highlighter"): This uses advanced photography terminology to frame the extreme optical phenomena rarely seen by the base model, guiding the AI to focus on and learn these specific rendering techniques.</li></ol>Dataset Categorization &amp; Training Focus: The dataset was divided into four specialized categories, each with a unique tagging focus:<ul><li>A. Single Person (Face/Full/Half body): Anchors character appearance and clothing details.<ul><li>WD: Character trigger words + physical details.</li><li>Qwen: Lighting rendering + skin materials (subsurface scattering).</li></ul></li><li>B. Multiple People (Interaction): Anchors composition, spatial relations, and positioning.<ul><li>WD: Interaction actions + camera angles.</li><li>Qwen: Depth of field, perspective, and 3D spatial awareness.</li></ul></li><li>C. Scenery (No humans): Anchors the overall style and atmospheric environment.<ul><li>WD: Geographical concept words.</li><li>Qwen: Macro composition, weather/climate, and spatial depth.</li></ul></li><li>D. Close-up (Objects/Hands/Feet): Anchors micro-materials and complex structures.<ul><li>WD: Object names.</li><li>Qwen: Micro-textures, metal/fabric textures, and focal expression.</li></ul></li></ul>Core Vocabulary: The core lighting and shadow vocabulary for the entire tagging scheme consists of about 15 high-frequency standard terms, tailored to the 4 folders to form a technical natural language tag system for LoRA style training:<ul><li>Lighting: <code>volumetric lighting</code>, <code>soft bloom</code>, <code>light diffusion</code>, <code>specular highlights</code></li><li>Materials: <code>translucent shading</code>, <code>subsurface scattering</code>, <code>material response</code>, <code>surface texture</code></li><li>Spatial: <code>depth layering</code>, <code>atmospheric perspective</code>, <code>feathered edges</code></li><li>Techniques: <code>painterly rendering</code>, <code>layered color washes</code>, <code>soft gradient blending</code></li></ul>Developer's Note: Because this hybrid tagging is a first-time experiment, the final output might not perfectly match my highest expectations yet. I will continue to explore and refine this methodology in the future!紫罗兰 LoRA V2 版本发布说明触发词： <code>violet cinematography</code>核心更新： 本次 V2 版本成功将光影效果和材质效果进行了剥离，使得它们不再与触发词强关联，避免了所有画面都被强制渲染出过于浓郁的“紫罗兰味”。但是！在增加画面可控性的同时，这也相应地提高了提示词（Prompt）的编写难度。训练集扩充： 本次大幅扩大了训练集，完整收录了十二集番剧和两部剧场版，训练集总计高达 6234 张剧照。若是以后修改训练方向或者进一步优化训练集，恐怕 6234 这个数量就已经是上限了。全新打标逻辑： 本次初次采用了 WD14 单词打标 + 千问自然语言混合打标 的方案：<ol><li>WD14 单词标的作用是“挡枪”： 把我们不需要模型学的东西（特定的衣服、动作、长相）用单词解释掉，告诉模型“这部分你底模自己懂，别把它们算进我的新画风里”。</li><li>千问自然语言的作用是“点睛”： 用高阶的摄影术语，把底模平时很少见到的“极品光学现象”框选出来，告诉她“好好看，好好学”。</li></ol>分类类型与打标侧重： 本次训练集被精细划分为四大类，分别对应不同的训练目的：<ul><li>A. 单人图 (人脸+全/半身)：<ul><li>主要目的： 锚定角色长相、服装细节。</li><li>打标侧重： WD 负责角色触发词+物理细节；千问负责光影渲染+皮肤材质。</li></ul></li><li>B. 多人交互图：<ul><li>主要目的： 锚定构图、空间关系、站位。</li><li>打标侧重： WD 负责交互动作、机位标签；千问负责景深、透视、三维空间感。</li></ul></li><li>C. 大场景 (无人)：<ul><li>主要目的： 锚定画风、空间氛围。</li><li>打标侧重： WD 负责地理概念词；千问负责宏观构图、气候气象、空间纵深。</li></ul></li><li>D. 特写 (物品/手脚)：<ul><li>主要目的： 锚定微观材质、复杂结构。</li><li>打标侧重： WD 负责物品名称；千问负责微观纹理、金属/布料质感、焦距表现。</li></ul></li></ul>核心光影词汇集： 整个打标方案的核心光影词汇集约为 15 个高频标准术语，配合 4 个文件夹各自的特化方向，构成了一套面向 LoRA 画风训练的技术性自然语言标签体系：<ul><li>光照类： <code>volumetric lighting</code>, <code>soft bloom</code>, <code>light diffusion</code>, <code>specular highlights</code></li><li>材质类： <code>translucent shading</code>, <code>subsurface scattering</code>, <code>material response</code>, <code>surface texture</code></li><li>空间类： <code>depth layering</code>, <code>atmospheric perspective</code>, <code>feathered edges</code></li><li>绘画技法类： <code>painterly rendering</code>, <code>layered color washes</code>, <code>soft gradient blending</code></li></ul>开发者寄语：因为是初次进行这种混合打标实验，所以最终效果可能没有想象中那么完美无缺，只能说日后会继续多多探索与优化！

Files with SHA256: dd2439fabc7dcb54b8adae498ed6cc1903d7d3a6ab8a281b982fdd4d7005f1bb