<h1 id="krea2-east-asian-portrait-refiner">Krea2 East Asian Portrait Refiner</h1><h2>中文介绍</h2>Krea2 East Asian Portrait Refiner 是一个针对 Krea 2 的亚洲人物面部与人像质感增强 LoRA。它不是单人物角色 LoRA，也不是某一种固定“网红脸”或审美风格 LoRA。训练集包含大量不同人物，同时包含较多古风、汉服以及高质量人物摄影。这个版本主要针对 Face / 面部表现 进行训练和数据处理，希望改善 Krea 2 在亚洲人物生成中比较容易出现的几个问题：<ul><li>亚洲面部变化不够丰富</li><li>多人物训练后容易逐渐出现相似面孔</li><li>部分亚洲人物皮肤和五官质感偏油画、偏绘画</li><li>真实摄影中的面部细节和皮肤质感不足</li><li>古风、汉服场景中人物面部容易被整体风格弱化</li><li>不同脸型之间的区分度不足</li></ul>这个 LoRA 不需要专用触发词，正常加载后直接使用自然语言提示词即可。<hr /><h2>关于平均脸</h2>减少多人物训练中的“平均脸 / 同质化”是这次训练最重要的实验目标之一。训练前，我使用人脸识别特征对数据中的人物进行了保守划分，希望让不同面孔尽可能保持独立的训练模式，而不是把所有人物直接混在一起。大致流程为：<code>Images → Face Detection → Face Embedding → Identity Grouping → Dataset Filtering → LoRA Training</code>这种处理目前确实能够部分缓解不同人物向同一张脸收敛的问题，但并没有彻底解决。部分提示词、随机种子或者较高 LoRA 强度下，仍然可以观察到不同人物出现一定程度的共同面部特征。所以这一版本并不宣称“解决了平均脸”。更准确地说，它是：一次利用 Face Identity 划分来降低多人物 LoRA 面部互相平均的实验。这一部分仍然在继续调整。如果你有多人物 LoRA、身份解耦、采样均衡、Caption、LoRA Rank 或 Target Layer 方面的经验，非常欢迎提供建议。尤其欢迎关于：<ul><li>不同人物样本数量不平衡</li><li>少样本人脸学习不足</li><li>多身份共享 LoRA 参数导致的干扰</li><li>Face token / Caption 设计</li><li>LoRA target layers</li><li>Rank 与容量</li><li>人物采样策略</li></ul>这些问题的反馈。<hr /><h2>数据集</h2>训练数据以不同亚洲人物为主体，其中包含较多：<ul><li>现代人物摄影</li><li>古风人物摄影</li><li>汉服人物摄影</li><li>室内人像</li><li>户外自然光摄影</li><li>不同脸型、五官和妆容</li><li>正面、侧面和不同头部角度</li><li>半身、近景以及部分全身人物</li></ul>其中汉服和古风图片占有一定比例。不过需要特别说明：这个版本只针对面部进行了专门的数据划分。目前没有对：<ul><li>汉服朝代</li><li>具体服饰制度</li><li>身材</li><li>姿势</li><li>服装类型</li></ul>进行专门的聚类监督。因此它不是一个“明制 LoRA”或者“汉服分类 LoRA”。训练数据只是包含了较丰富的传统中式服装和古风人物场景。在实际测试中，Krea 2 原有的服装语义能力仍然能够正常工作，因此可以直接在提示词中指定具体服装款式，例如：<code>mamian skirt</code><code>standing collar</code><code>wide sleeves</code><code>cross-collar hanfu</code><code>traditional Chinese clothing</code>具体服装仍然主要依靠自然语言提示词控制，而不是依赖本 LoRA 内部的朝代或款式标签。包含部分NSFW数据，是为了避免抑制NSFW生成，但并非为了生成NSFW内容。<hr /><h2>主要效果</h2>这个 LoRA 主要希望改善：<ul><li>亚洲人物面部多样性</li><li>不同人物之间的脸型区分</li><li>面部摄影质感</li><li>自然皮肤纹理</li><li>眼睛、鼻部、嘴唇等面部细节</li><li>亚洲人物容易出现的偏油画质感</li><li>古风和汉服摄影中的人物面部表现</li><li>高质量摄影环境下的人像质感</li></ul>它更接近一个：East Asian Portrait / Face Enhancement LoRA而不是角色 LoRA。<hr /><h2>使用方法</h2>无需 Trigger Word。正常加载 LoRA 后直接写提示词即可。例如：<code>Chinese woman, realistic portrait photography, natural skin texture, soft daylight</code>或者：<code>Chinese woman wearing hanfu, traditional Chinese clothing, natural light, realistic photography</code>需要具体汉服款式时，可以直接继续描述服装。推荐先从常规 LoRA 权重开始，再根据工作流和希望的强化程度调整。不同工作流下最佳权重可能有所区别。<hr /><h2>当前状态</h2>这是一个实验版本。当前比较确定的改善主要集中在：亚洲人物面部表现与摄影质感。Face Identity 分组对于平均脸问题有一定帮助，但目前仍然只能算部分缓解。我会继续尝试调整：<ul><li>人物样本均衡</li><li>Face Identity 条件</li><li>LoRA target layers</li><li>Rank</li><li>Caption</li><li>训练数据分布</li></ul>如果你发现：<ul><li>某些提示词特别容易出现平均脸</li><li>某些权重下脸型开始明显同质化</li><li>汉服或现代人物中的面部表现差异</li><li>特定构图下效果明显变差</li></ul>欢迎在评论区反馈。<hr /><h1 id="english-description">English Description</h1>Krea2 East Asian Portrait Refiner is a facial and portrait enhancement LoRA for Krea 2, focused primarily on East Asian subjects.This is not a single-character LoRA and it is not designed to reproduce one fixed facial style or one particular beauty standard.The dataset contains many different identities, together with a substantial amount of high-quality modern portrait photography, traditional Chinese-style photography, and Hanfu imagery.The main focus of this version is facial representation.It attempts to improve several issues that can appear when generating East Asian subjects with Krea 2:<ul><li>limited variation between Asian facial appearances</li><li>different subjects gradually drifting toward similar faces</li><li>overly painterly facial or skin texture</li><li>insufficient photographic skin detail</li><li>weaker facial definition in traditional Chinese / Hanfu scenes</li><li>limited separation between different facial structures</li></ul>No dedicated trigger word is required.Simply load the LoRA and use normal prompts.<hr /><h2 id="about-the-average-face-problem">About the Average-Face Problem</h2>Reducing average-face / same-face behavior in multi-person training was one of the main experimental goals of this project.Before training, face-recognition embeddings were used to conservatively separate different identities.The simplified preprocessing pipeline is:<code>Images → Face Detection → Face Embedding → Identity Grouping → Dataset Filtering → LoRA Training</code>The idea is to avoid treating hundreds of visually different people as one completely homogeneous training distribution.This approach does appear to partially reduce facial averaging, but it does not fully solve the problem.With some prompts, seeds, or higher LoRA strengths, different subjects can still drift toward similar facial characteristics.So this model should not be described as a complete solution to multi-identity face collapse.It is more accurately:an experiment in using face-identity separation to reduce interference between different facial modes inside one LoRA.Feedback is very welcome, especially from people with experience in:<ul><li>multi-identity LoRA training</li><li>identity disentanglement</li><li>imbalanced identity datasets</li><li>sampling strategies</li><li>caption design</li><li>LoRA target layers</li><li>rank / capacity selection</li></ul>I am particularly interested in better ways to preserve multiple distinct facial modes inside a single LoRA without allowing them to gradually converge toward a shared appearance.<hr /><h2 id="dataset">Dataset</h2>The training dataset contains many different East Asian subjects and includes:<ul><li>modern portrait photography</li><li>traditional Chinese-style photography</li><li>Hanfu photography</li><li>indoor portraits</li><li>outdoor natural-light photography</li><li>different facial structures and makeup styles</li><li>frontal, three-quarter and profile views</li><li>close-up, upper-body and some full-body compositions</li></ul>A noticeable portion of the dataset contains Hanfu and traditional Chinese aesthetics.However:only facial identity received specialized preprocessing in this version.There is currently no dedicated training-label system for:<ul><li>dynasty</li><li>historical clothing system</li><li>body shape</li><li>pose</li><li>clothing categories</li></ul>Therefore this should not be considered a dynasty-specific or dedicated Hanfu classification LoRA.The dataset simply contains a relatively rich amount of traditional Chinese clothing and historical-style portrait imagery.Krea 2's original language understanding for clothing remains useful, so specific garment concepts can still be prompted directly, for example:<code>mamian skirt</code><code>standing collar</code><code>wide sleeves</code><code>cross-collar hanfu</code><code>traditional Chinese clothing</code>Specific garments are primarily controlled through natural-language prompts rather than internal dynasty or clothing labels.Contains some NSFW data to prevent the suppression of NSFW generation, but not for the purpose of generating NSFW content.<hr /><h2 id="main-goals">Main Goals</h2>The LoRA primarily aims to improve:<ul><li>diversity of East Asian facial appearances</li><li>separation between different facial structures</li><li>realistic portrait texture</li><li>natural skin rendering</li><li>facial detail</li><li>reduced painterly facial appearance</li><li>facial quality in Hanfu and traditional Chinese portrait scenes</li><li>overall photographic rendering of East Asian subjects</li></ul>It is best understood as an:East Asian Portrait / Face Enhancement LoRArather than a character LoRA.<hr /><h2 id="usage">Usage</h2>No Trigger Word Required.Load the LoRA and prompt normally.For example:<code>Chinese woman, realistic portrait photography, natural skin texture, soft daylight</code>or:<code>Chinese woman wearing hanfu, traditional Chinese clothing, natural light, realistic photography</code>Specific clothing styles can be described directly in the prompt.Start with a normal LoRA strength and adjust it according to your workflow and desired effect.<hr /><h2 id="current-status">Current Status</h2>This is still an experimental release.The most consistent improvements currently appear in:East Asian facial rendering and photographic portrait quality.Face identity grouping appears to partially reduce average-face behavior, but it does not eliminate it.Future experiments will continue to explore:<ul><li>identity-balanced sampling</li><li>facial identity conditioning</li><li>LoRA target layers</li><li>rank and capacity</li><li>caption strategies</li><li>dataset composition</li></ul>If you notice specific prompts, seeds, compositions, or LoRA strengths that make the average-face issue more obvious, feedback is very welcome.

Files with SHA256: 2e1dbb2e066b8ab83fbaf5f912139ad681ea7ef2b00ab3524f830514bff18fbc