<p>runninghub cn:<a target="_blank" rel="ugc" href="https://www.runninghub.cn/ai-detail/2095217490568761345">https://www.runninghub.cn/ai-detail/2095217490568761345</a></p><p>runninghub ai:<a target="_blank" rel="ugc" href="https://www.runninghub.ai/ai-detail/2093900524172447746">https://www.runninghub.ai/ai-detail/2093900524172447746</a></p><p>huggingface:<a target="_blank" rel="ugc" href="https://huggingface.co/Jojocodex/wushu-action-v7-minimax-h3-fl2va-ref2va-lora">https://huggingface.co/Jojocodex/wushu-action-v7-minimax-h3-fl2va-ref2va-lora</a></p><h2><span><strong>使用前必读</strong></span></h2><ol><li><p><strong>触发词 </strong><code>wushu_action</code><strong> 必须放在提示词最前面</strong>。所有训练标注均以此词开头，模型将该词与武术动作强绑定。</p></li><li><p>该LoRA训练素材全部为人体模型三视图动画，默认容易输出白模+纯黑背景。这是设计特性，不是BUG。</p></li><li><p>LoRA权重推荐范围：</p><ul><li><p>0.9–1.0：生成三视图人体动作参考素材</p></li><li><p><strong>0.6–0.8</strong>：正常场景出片（推荐起点）</p></li><li><p>若大量出现白模黑底，降低至0.5–0.6</p></li></ul></li></ol><h2><span><strong>提示词写法</strong></span></h2><h3 id="a"><span><strong>模式A：生成训练风格三视图人体动作参考</strong></span></h3><p>直接复制模板，仅替换招式描述</p><pre><code>wushu_action, 白色无贴图男性人体模型, 同一角色三视图同步并排, 纯黑虚空, 固定全身机位, 动作完全同步, 不是三个战士,【填入招式】, 实时爆发力, 全身发力传导, 无慢动作, 无停顿
</code></pre><h3 id="b"><span><strong>模式B：生成真实角色与场景（视频创作推荐）</strong></span></h3><p>保留触发词和动作关键词，把模板里的人体模型、三视图、黑底替换为你需要的角色和场景</p><pre><code>wushu_action,【角色与场景描述】,【填入招式】, 实时爆发力, 全身发力传导, 无慢动作, 无停顿
</code></pre><blockquote><p><strong><em>提示：优先使用训练集中已存在招式词，成功率更高：直拳、高侧踢、疾步冲刺、举盾格挡、长枪突刺、翻滚、连续连招</em></strong></p></blockquote><h2 id="h3"><span><strong>H3 生成参数</strong></span></h2><ul><li><p>采样器：euler｜调度器：simple</p></li><li><p>采样步数：25</p></li><li><p>CFG：<strong>1.0，请勿大于1</strong></p></li><li><p>视频帧数：遵守 <code>17n+5</code> 规则（可选：22 / 39 / 73 / 90 / 124）</p></li><li><p>分辨率：宽高必须为32的倍数，推荐 832×480</p></li></ul><h2><span><strong>负面提示词</strong></span></h2><pre><code>模糊，扭曲，低画质，画面抖动，多余肢体，手部畸形，形体变形，光影错乱，画面闪烁，静止姿势，动作僵硬
</code></pre><h2><span><strong>已知限制</strong></span></h2><ul><li><p>训练片段时长仅1.3–3秒。长视频建议分段生成后剪辑拼接。</p></li><li><p>训练素材均为单人，双人对打属于外推效果，稳定性较差。</p></li><li><p>本LoRA只学习动作；角色皮肤、服装、面部质感依靠底模或画风LoRA实现。</p></li></ul><h2 id="comfyui"><span><strong>ComfyUI部署方法</strong></span></h2><ol><li><p>将LoRA文件放入 <code>models/loras/</code> 文件夹</p></li><li><p>加载 <code>minimax_h3_fl2va</code> 扩散模型，搭配对应的Qwen3VL文本编码器、H3视频VAE</p></li><li><p>UNet加载节点后接入LoRA加载器，设置上面推荐的权重</p></li><li><p>LoRA挂载校验：对比权重0与权重1的出片效果。若无明显差别，需要对safetensors做键名重映射。</p></li></ol><p>Wushu_Action LoRA Official Usage Guide</p><h2 id="mandatory-prefix-rule">Mandatory Prefix Rule</h2><p>The trigger word <strong>wushu_actionmust be placed at the very beginning</strong> of every prompt. All training data is anchored to this prefix, which strongly binds the model to recognize and generate authentic martial arts movements.</p><h2 id="model-characteristics">Model Characteristics</h2><p>This LoRA is trained exclusively on <strong>3-view orthographic human animation references</strong> (front/side/back). Default outputs tend to feature <strong>white textureless mannequins with pure black void backgrounds</strong>. This is an intended design feature, not a bug.</p><h2 id="recommended-lora-weight-range">Recommended LoRA Weight Range</h2><ul><li><p><strong>0.9 – 1.0</strong>: For generating standard 3-view human motion reference assets</p></li><li><p><strong>0.6 – 0.8</strong>: General scene video generation (best starting point for most creations)</p></li><li><p><strong>0.5 – 0.6</strong>: Use if white mannequins or black empty backgrounds appear excessively</p></li></ul><h2 id="prompt-writing-templates">Prompt Writing Templates</h2><h3 id="mode-a:-3-view-orthographic-motion-reference-(training-style)">Mode A: 3-View Orthographic Motion Reference (Training Style)</h3><p>Use the full template, only replace the martial arts move description:</p><p><strong>wushu_action</strong>, white textureless male human mannequin, same character displayed in synchronized 3-view layout, pure black void background, fixed full-body camera, perfectly synchronized movement, no three separate characters, <strong>[insert martial arts move]</strong>, real-time explosive power, full-body force transmission, no slow motion, no pause</p><h3 id="mode-b:-realistic-character-and-scene-video-(recommended-for-video-production)">Mode B: Realistic Character &amp; Scene Video (Recommended for Video Production)</h3><p>Keep the trigger word and motion keywords; replace mannequin/3-view/black background terms with your custom characters and scenes:</p><p><strong>wushu_action</strong>, <strong>[character &amp; scene description]</strong>, <strong>[insert martial arts move]</strong>, real-time explosive power, full-body force transmission, no slow motion, no pause</p><h3 id="high-success-pre-trained-motion-vocabulary">High-Success Pre-Trained Motion Vocabulary</h3><p>These movements exist in the training dataset and yield the most stable results: Straight punch, high side kick, rapid dash, shield block, spear thrust, roll dodge, continuous combo attacks</p><h2 id="h3-generation-parameters">H3 Generation Parameters</h2><ul><li><p>Sampler: Euler</p></li><li><p>Scheduler: Simple</p></li><li><p>Sampling Steps: 25</p></li><li><p>CFG Scale: 1.0 (<strong>do not exceed 1.0</strong>)</p></li><li><p>Video Frame Rule: Follow <strong>17n+5</strong> sequence (Recommended: 22 / 39 / 73 / 90 / 124 frames)</p></li><li><p>Resolution: Width &amp; Height must be multiples of 32. Recommended: 832×480</p></li></ul><h2 id="negative-prompt">Negative Prompt</h2><p>blurry, distorted, low quality, screen jitter, extra limbs, hand deformity, body distortion, wrong lighting, flickering screen, static pose, stiff movement</p><h2 id="known-limitations">Known Limitations</h2><ul><li><p>Original training clips are only <strong>1.3–3 seconds</strong> long. For long videos, generate segmented clips and edit them together.</p></li><li><p>All training data features <strong>single-person movements</strong>. Dual-person combat is an extrapolated effect with lower stability.</p></li><li><p>This LoRA only learns <strong>motion logic</strong>. Character skin, clothing, facial details and rendering styles rely on the base model or style LoRAs.</p></li></ul><h2 id="comfyui-deployment-guide">ComfyUI Deployment Guide</h2><ol><li><p>Place the LoRA file into: <code>models/loras/</code></p></li><li><p>Load the <strong>minimax_h3_fl2va</strong> diffusion model, paired with the matching Qwen3VL text encoder and H3 Video VAE</p></li><li><p>Connect the LoRA loader node to the UNet node and apply the recommended weight values</p></li><li><p>LoRA Validation Test: Compare outputs at weight 0 and weight 1. If no visible motion difference exists, remap the safetensors key names manually.</p></li></ol><p></p>

Files with SHA256: 5f81a8f20255d7136eb6ea445b509aae40a975169097f745cd749264d3d788dc