About this LoRAI noticed that Anima often struggle to accurately interpret concepts like "tall" or "long legs". To solve this, I created this LoRA to help you easily generate tall (or even exceptionally tall) female characters.How it WorksPlease note: This LoRA does not simply "scale up" or magnify the character (e.g., making them look like a giant against the background). Instead, it authentically conveys the "tall" trait by physically adjusting the character's anatomical proportions—specifically by increasing the head-to-body ratio and elongating the legs.关于这个 LoRA我注意到Anima通常对“高个子”或“大长腿”等概念的理解不够理想，因此我制作了这个 LoRA。它可以帮助你生成身材高挑（甚至是非常高）的女性角色。模型特点需要特别说明的是：它并非简单地将人物“放大”（例如通过微缩场景来衬托巨大的角色），而是通过真实改变人物的比例（增大头身比、拉长双腿）来展现“高挑”的视觉特征。

Super Tall Girl with Long Legs/长腿高挑女孩

<s>重要警告：如果你在ComfyUI中使用，推荐模型权重约为 </s><code>0.04</code><s>。请从 </s><code>0.03</code><s> 开始调整，不要按照普通 LoRA 的习惯使用 </s><code>1.0</code><s>，否则会发生严重的权重坍缩。</s><s>IMPORTANT WARNING:If you are using ComfyUI, the recommended model strength is around </s><code>0.04</code><s>. Start at </s><code>0.03</code><s>. Do not use the usual LoRA strength of </s><code>1.0</code><s>, as it will cause severe weight collapse.</s> 我重新转换了模型，现在它可以按照正常lora的工作范围运行（虽然我依然觉得加载器没有正确应用缩放是一个bug) I re-converted the model, and now it can operate within the normal LoRA range (although I still think the loader not applying the scaling correctly is a bug) 由于ComfyUI的一个漏洞，LoKr缩放没有被正确应用，导致权重放大32倍，我已经提交至github，也许未来的某一天ComfyUI会修复此漏洞。 Due to a bug in ComfyUI, LoKr scaling wasn't applied correctly, causing the weights to be amplified 32 times. I've submitted it to GitHub, maybe one day ComfyUI will fix this bug. <a target="_blank" rel="ugc" href="https://github.com/Comfy-Org/ComfyUI/issues/15114">LoKr alpha scaling is ignored when both lokr_w1 and lokr_w2 are stored directly · Issue #15114 · Comfy-Org/ComfyUI</a> 模型简介<code>xtall v2</code> 是一个 LoKr 模型，用于控制超高头身比、修长体型和长腿表现。相比 v1，v2 更强调“按需生效”：没有身材相关提示词时，它通常不会主动生成夸张长腿；明确加入并强化相关概念后，则可以达到比 v1 更极端的身体比例，同时保留更好的画面稳定性。推荐权重<ul><li>推荐起点：<code>1</code></li><li>常用范围：<code>0.9-1.6</code></li><li>强效果范围：<code>1.6-2.2</code></li><li>极限边界：<code>2.8</code></li></ul>蒸馏模型通常需要更低的权重。完整模型可以尝试稍高权重，但甜点区仍然在 <code>1.1</code> 左右。触发词与用法<ul><li>基础触发词：<code>xtall</code></li><li>长腿：<code>long legs</code></li></ul>只使用 <code>xtall</code> 时，模型更倾向于整体修长和较高头身比。加入并提高 <code>long legs</code> 的权重后（甚至可以提升至10，仍然可以获得正常的图片），会着重强调长腿。如果只希望获得高头身比，请不要加入 <code>long legs</code>，并将模型权重保持在 <code>0.9-1.2</code>。模型特点<ul><li>比 v1 更低的中性提示词泄漏</li><li>对长腿强度具有更清晰的可调范围</li><li>比 v1 更高的极端身材上限</li><li>更小的画风污染</li><li>更好的自然语言理解能力（存疑）</li><li>更强的姿势泛化能力</li></ul>已知问题<ul><li>导致角色倾向于丰满，如果你不希望角色丰满，需要特别用提示词</li><li>增加头身比的同时，可能导致角色头也一起变大（与画面中其它人相比） </li></ul>Model Description<code>xtall v2</code> is a LoKr model designed to control very high head-to-body proportions, elongated body shapes, and long legs.Compared with v1, v2 is more prompt-dependent: without body-proportion-related prompts, it generally does not force exaggeratedly long legs. When the relevant concepts are explicitly added and strengthened, it can achieve more extreme proportions than v1 while maintaining better image stability.Recommended Model Strength<ul><li>Recommended starting point: <code>1</code></li><li>Normal range: <code>0.9-1.6</code></li><li>Strong-effect range: <code>1.6-2.2</code></li><li>Upper limit: <code>2.8</code></li></ul>Distilled models usually require a lower strength. Full models may tolerate slightly higher values, but the sweet spot is still around <code>1.1</code>.Trigger Words and Usage<ul><li>Main trigger: <code>xtall</code></li><li>Long legs: <code>long legs</code></li></ul>Using <code>xtall</code> alone tends to produce an overall taller, more elongated figure with a higher head-to-body proportion. Adding <code>long legs</code> and increasing its prompt weight emphasizes leg length. The prompt weight for <code>long legs</code> can even be raised to <code>10</code> while still producing a coherent image.If you only want a high head-to-body proportion without strongly emphasizing the legs, omit <code>long legs</code> and keep the model strength around <code>0.9-1.2</code>.Model Features<ul><li>Less neutral-prompt leakage than v1</li><li>A clearer and more controllable response to long-leg strength</li><li>A higher ceiling for extreme body proportions than v1</li><li>Less interference with the original art style</li><li>Better natural-language understanding, though this remains uncertain</li><li>Stronger pose generalization</li></ul>Known Issues<ul><li>The model tends to make characters fuller or more voluptuous. If this is not desired, you may need to explicitly counteract it in the prompt.</li><li>While increasing the head-to-body proportion, the model may also enlarge the character’s head, especially relative to other people in the same image.</li></ul>

Files with SHA256: 3b17095d9957f89543d2929ce45baf3c4c689e5ddefe083f1ec09d2fa5032796