✨ 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 Words
为了保证最纯净的特征调用，本次 V3 版本的人物触发词已全面摒弃下划线格式，直接使用自然语言英文真名即可稳定召唤。 To ensure the purest feature activation, character triggers in V3 have completely abandoned the underscore format. Simply use their natural language English names for stable summoning.
🎨 画风核心触发词 / Core Style Trigger:
violet cinematography(建议置于正向提示词首位 / Highly recommended to place at the beginning of the prompt)👥 人物一键召唤 / Character Triggers:
薇尔莉特·伊芙加登:
Violet Evergarden嘉德丽雅·波德莱尔:
Cattleya Baudelaire爱丽丝·卡娜莉:
Iris Cannary艾丽卡·布朗:
Erica Brown克劳迪亚·霍金斯:
Claudia Hodgins贝内迪克特·布卢:
Benedict Blue
💡 召唤建议 / Summoning Tip: 推荐使用“触发词 + 基础外貌描述 + 服饰描述 + 场景”的自然语言句式。人物触发词在训练集打标中已通过底层逻辑隔离，召唤极其稳定且绝不会污染整体画风。 We recommend using natural language prompts like "Trigger word + basic appearance + clothing + background." Character triggers are logically isolated during dataset tagging, ensuring stable summoning without polluting the overall art style.
关于 V3 终极形态：全域重构与技术升级 / About V3 Ultimate Form: Full-Domain Reconstruction & Tech Upgrade
V3 不是一次简单的版本迭代，而是一场从底层逻辑开始的全面重构。在出现跨时代的全新训练理论之前，本套训练集不会再有大规模变动。 V3 is not a simple version iteration, but a comprehensive reconstruction from the underlying logic. Unless there are completely innovative training theories in the future, this dataset will not undergo massive changes again.
1. 彻底治愈“半身像依赖症” / Curing the "Half-Body Portrait Bias"
[CN] 旧版（V2）由于半身像占主导，导致生成时模型犹如一位“顽固的摄影师”，极易偏科。V3 通过极其严苛的分类与留存策略，保留了从特写、半身、全身到宏大场景的科学比例。现在，你可以通过提示词自由且精准地控制画面景别，再也不会被强制输出半身像。
[EN] In previous versions, half-body shots dominated the dataset, causing the AI to aggressively generate them regardless of your prompt. V3 implements strict categorization and retention strategies to maintain a scientific ratio of close-ups, half-bodies, full-bodies, and grand scenes. Now, you can freely and precisely control the shot composition without being forced into half-body outputs.
2. 工业级五步筛选流 / Industrial-Grade 5-Step Sorting Pipeline
[CN] 训练集原始素材超 10,000 张，经过以下五步硬核提纯，最终精选约 3000 张极品素材：
[EN] The original dataset of over 10,000 images was rigorously purified into approximately 3,000 elite assets through these five steps:
CLIP 亮度初筛 (CLIP Brightness Sorting): 剥离过曝与死黑废片。 / Stripping away overexposed and pitch-black unusable images.
千问 7B 语义分类 (Qwen 7B Semantic Sorting): 粗分单人、多人、场景与特写。 / Roughly categorizing into single, multiple people, scenes, and close-ups.
人工像素级精检 (Manual Pixel-Level Curation): 借助 FastStone Image Viewer 与 DigiKam 进行人工纠错微调。 / Manual error correction and fine-tuning using image viewers.
WD14 基础特征提取 (WD14 Feature Extraction): 提取客观存在的物理元素标签。 / Extracting objective physical element tags.
千问 7B 视觉注入 (Qwen 7B Visual Injection): 追加自然语言的光影与材质描述。 / Appending natural language descriptions for lighting and materials.
3. 差异化分层训练 / Differentiated Stratified Training
[CN] 本次全面迁移至高阶脚本训练器，摒弃“一视同仁”的做法。我们根据四大核心分类，按文件夹精细分配了不同的重复遍数（Repeats），彻底解决过拟合与欠拟合风险：
[EN] Fully migrated to advanced training scripts, abandoning the "treat all equally" approach. We assigned specific repeat counts based on four core folders to eliminate over/under-fitting risks:
👤 单人图 (Single Person): 锚定角色长相与服装。WD14负责物理特征，千问负责光影与皮肤材质(SSS)。 / Anchors character looks. WD14 handles physical traits; Qwen handles lighting and skin subsurface scattering.
👥 多人交互 (Multiple People): 锚定构图与空间关系。WD14负责动作与机位，千问负责景深与三维透视。 / Anchors composition. WD14 handles actions; Qwen handles depth of field and 3D perspective.
🏔️ 大场景 (Scenery): 锚定画风基调与空间氛围。WD14负责基础地理概念，千问负责宏观构图、气象与空间纵深。 / Anchors the core style. WD14 handles geography; Qwen handles macro composition and atmosphere.
🔍 特写 (Close-ups): 锚定微观材质。WD14负责物品名称，千问负责微观纹理、金属/布料质感及焦距表现。 / Anchors micro-materials. WD14 names objects; Qwen handles textures, metallic/fabric feel, and focus.
4. 核心打标体系：WD14 + 千问自然语言双擎驱动 / WD14 + Qwen Hybrid Tagging Engine
[CN] 彻底抛弃纯单词堆砌，采用“底层词汇挡枪 + 自然语言点睛”的双层混合架构：
[EN] Completely abandoning pure word-stacking, we adopted a dual-layer hybrid architecture: "Base tags as a shield + Natural language as the finishing touch."
🛡️ WD14 标签 (防守 / Defense): 负责描述“画了什么”。告诉模型“这些动作和衣服你本来就懂，不用把它们算进我的新画风里”。 / Describes "what is drawn," telling the model not to bake basic poses and clothes into the new style.
⚔️ 千问 7B 标签 (进攻 / Offense): 负责描述“怎么画的”。强制引导模型去学习平时极难见到的“极品光学现象”。 / Describes "how it's drawn," forcing the model to learn rare, top-tier optical phenomena.
5. 核心光影词汇集 / Core Lighting Vocabulary
[CN] 为确保画风高度一致，自然语言打标统一使用了 15 个高频专业摄影术语：
[EN] To ensure style consistency, the natural language tagging utilizes 15 high-frequency professional photography terms:
光照类 (Lighting): volumetric lighting, soft bloom, light diffusion, specular highlights
材质类 (Materials): translucent shading, subsurface scattering, material response, surface texture
空间类 (Spatial): depth layering, atmospheric perspective, feathered edges
绘画技法 (Techniques): painterly rendering, layered color washes, soft gradient blending
📸 求返图！交流学习 / Share Your Art!
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!
⚠️ 免责声明 / Disclaimer
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.