<h3 id="general-motion-continuity-repair-lora">General Motion Continuity Repair LoRA</h3><h3 id="lora">通用动作连续性修复 LoRA</h3>This is a general-purpose motion support LoRA, not a combat-only LoRA. It can be used for running / sports / dance / acrobatics / character interaction / combat / weapon motion and other dynamic scenes. 这是一个通用动态辅助 LoRA，并不是战斗专用模型，可用于跑步 / 体育 / 舞蹈 / 翻滚跳跃 / 人物互动 / 战斗 / 武器动作等动态场景。<h3 id="if-the-lora-looks-muddy-or-less-impactful-than-expected-check-the-sampler-setup-first.">If the LoRA looks muddy or less impactful than expected, check the sampler setup first.</h3>Sampler choice matters a lot for this LoRA.For best clarity and motion definition, I recommend one of the following setups:Option A<ul><li>Sampler: <code>res_multistep</code></li><li>Scheduler:simple</li></ul>Option B<ul><li>Sampler: <code>euler</code></li><li>Scheduler: <code>beta</code></li></ul>Other sampler / scheduler combinations may produce noticeably blurrier or softer results.🚀 Run It on RunningHub🎁 New users on RunningHub International can get 1000 RH Coins after registration.<a target="_blank" rel="ugc" href="https://www.runninghub.ai/zh-cn/post/2093321887345819650/?inviteCode=lk4qa7rh ">▶ Start / End Frame One-Click Workflow</a><a target="_blank" rel="ugc" href="https://www.runninghub.ai/zh-cn/post/2093328847935950849/?inviteCode=lk4qa7rh ">▶ Multi-Reference One-Click Workflow</a><a target="_blank" rel="ugc" href="https://www.runninghub.ai/zh-cn/user-center/2092843565587107841/model?inviteCode=lk4qa7rh">📦Bunny Official RunningHub Model Page</a>Future FourBunny LoRAs will also be released and updated on RunningHub.<hr /><h3 id="highly-recommended-for-complex-action:-pair-this-lora-with-bunny-h3-conditioning-bridge-for-better-multi-character-logic-action-ownership-and-scene-stability.-bunny-h3-conditioning-bridge"><a target="_blank" rel="ugc" href="https://civarchive.com/models/2936635/bunny-h3-conditioning-bridge-or-multi-character-logic-and-scene-stability">Highly recommended for complex action: pair this LoRA with BUNNY H3 Conditioning Bridge for better multi-character logic, action ownership, and scene stability. 复杂动作场景强烈推荐搭配 BUNNY H3 Conditioning Bridge，可进一步改善多人互动逻辑、动作归属和场景稳定性。</a></h3><h3 id="minimax-h312gb-vramuniversal-fl2varef2va-base-model-+-low-sigma-combat-second-pass-workflow"><a target="_blank" rel="ugc" href="https://civarchive.com/models/2869434/minimax-h312gb-vramuniversal-fl2varef2va-base-model-low-sigma-combat-second-pass-workflow?modelVersionId=3241900">【MiniMax H3】12GB VRAM｜Universal FL2VA/REF2VA Base Model + Low-Sigma Combat Second-Pass Workflow</a></h3><hr /><h2 id="recommended-usage">Recommended Usage / 推荐使用</h2>Standalone use: around <code>0.9</code> is currently recommended. 单独使用时，目前推荐权重约为 <code>0.9</code>。When used alone, Motion Continuity Repair needs a relatively high weight to produce a clear effect on motion continuity / slow-motion tendency / missing transitions / prompt-following. 单独使用时，需要相对较高的权重，才能比较明显地影响动作连续性 / 慢动作倾向 / 动作断链 / 提示词遵循度。<a target="_blank" rel="ugc" href="https://civarchive.com/models/2853878/minimax-h3-combat-base-fight-motion-impact-drama-booster">With a Combat LoRA: Stage 1 around </a><code>0.5–0.7</code><a target="_blank" rel="ugc" href="https://civarchive.com/models/2853878/minimax-h3-combat-base-fight-motion-impact-drama-booster"> is recommended. 搭配 Combat LoRA 时，一采推荐约 </a><code>0.5–0.7</code><a target="_blank" rel="ugc" href="https://civarchive.com/models/2853878/minimax-h3-combat-base-fight-motion-impact-drama-booster">。</a>Because Combat already provides strong motion speed and impact, Motion Continuity Repair only needs a medium weight to correct continuity / coordination / action logic without unnecessarily taking over the motion. 因为 Combat 已经提供了较强的动作速度和打击感，此时 Motion Continuity Repair 只需要中等权重来修正动作衔接 / 协调性 / 动作逻辑，避免过度接管整个动态。More complex sequences can gradually increase the Stage 1 weight, but very high weights may begin to reshape the entire motion logic / affect visual style / or alter audio characteristics. 动作越复杂可以逐渐提高一采权重，但过高时可能开始重构整段动作逻辑 / 改变画面表现 / 影响声音特征。Stage 2: around <code>0.2–0.3</code> is recommended. 二采建议约 <code>0.2–0.3</code>。Higher Stage 2 weights may over-stabilize the sequence and reduce motion intensity / speed / impact. 二采权重过高可能会把动作收得太稳，从而削弱动作幅度 / 速度感 / 冲击力。最简单地说就是：<blockquote>单独使用：≈0.9 / 搭配 Combat：0.5–0.7 / 二采：0.2–0.3</blockquote>这个才是目前测试出来的正确推荐方式。Although it is a general-purpose LoRA, it works especially well together with a Combat LoRA. Combat provides speed and impact / Motion Continuity Repair focuses on continuity, coordination, prompt-following and interaction logic. 虽然它是一个通用 LoRA，但目前非常推荐与 Combat LoRA 搭配使用。Combat 负责速度与打击感 / Motion Continuity Repair 负责动作衔接、协调性、提示词遵循与交互逻辑。<h2 id="trigger-word">Trigger Word / 触发词</h2>Trigger word: <code>bunny_crisp_motion</code> — optional. It mainly serves as an explicit identifier for the LoRA and is not required for normal use. 触发词：<code>bunny_crisp_motion</code> —— 可选。它主要用于明确标识这个 LoRA，正常使用时并不是必须填写。You can test with or without the trigger word depending on your workflow and prompt. 你可以根据自己的工作流和 Prompt，自行测试是否加入触发词。<hr /><h2 id="what-does-it-do">What does it do? / 它主要做什么？</h2>Ever had a great clip ruined by less than one second of strange motion? A sudden slowdown / a missing action / an awkward limb trajectory / a broken interaction can easily make an otherwise usable clip unusable. 你是否遇到过这种情况：一段视频明明整体很好，却偏偏有不到 1 秒突然慢下来 / 漏掉一个动作 / 肢体轨迹变得诡异 / 交互突然断掉，最后整段素材因此报废？This LoRA is designed around repairing those moments rather than simply making every motion stronger. 这个 LoRA 的重点就是修复这些“突然掉链子”的瞬间，而不是粗暴地把所有动作统一增强。Current A/B testing shows improvements in unwanted slow-motion behavior / missing motion transitions / abnormal limb trajectories / action coordination / prompt-following / interaction continuity / and some mismatched motion-feedback events. 目前 A/B 测试中已经观察到对 异常慢动作 / 动作断链 / 诡异肢体轨迹 / 动作协调性 / 提示词遵循度 / 交互连续性 / 部分错误动作反馈 的改善。When H3 already handles a scene well, the difference may be very small / when the base model begins to struggle, the difference can become much more obvious. 当 H3 本来就很擅长某段动态时，开启前后的区别可能非常小 / 当底模开始掉速、断动作或失去逻辑时，差异反而会明显很多。<blockquote>It doesn't simply make H3's strengths stronger — it tries to keep its weaker motion states from falling apart. 它不是单纯让 H3 的强项更强，而是尽量不让 H3 在弱项动态里突然掉下去。</blockquote><hr /><h2 id="current-ab-results">Current A/B Results / 当前对比表现</h2>Different weights can progressively change motion intensity and eventually begin reshaping the overall motion rhythm. 随着一采权重提高，可以逐步改变原本偏弱的动态表现，高权重下甚至会开始重构整体动作节奏。In complex action sequences, the LoRA can reconnect missing actions and reduce strange limb paths, making the motion chain more complete. 在复杂连续动作中，它可以补回部分没有连接上的动作，并减少诡异肢体轨迹，让动作链更加完整。In scenes where even a Combat LoRA cannot fully suppress H3's slow-motion tendency, Motion Continuity Repair can further pull the sequence back toward a faster real-time motion state. 在一些连 Combat LoRA 都无法完全压住慢动作倾向的场景中，Motion Continuity Repair 还能进一步把动作往正常高速节奏拉回。It can also improve action-prompt adherence, turning sequences that originally move randomly into results that follow the intended action order more closely. 它也可以提升动作提示词遵循度，让原本只是胡乱运动的结果，更容易按照 Prompt 中指定的动作顺序执行。<hr /><h2 id="weapon-interaction">Weapon Interaction / 武器交互</h2>A portion of weapon finisher / execution animation data was included in the dataset. 训练数据中还加入了一部分武器处决 / 终结动画。Current testing shows some improvement in weapon disappearance / clipping / unstable fast swings / incomplete hit reactions / and certain finishing or kill animations. 目前测试中，对 武器突然消失 / 穿模 / 高速挥动不稳定 / 命中反馈不完整 / 部分终结与击杀动作 都有一定改善。It is not a dedicated Weapon Combat LoRA, so sudden hand switching / unstable grips / occasional weapon disappearance / complex collision errors may still occur. 不过它不是专门的 Weapon Combat LoRA，因此突然换手 / 握持跳变 / 偶尔消失 / 复杂碰撞错误仍然可能发生。<hr /><h2 id="how-it-was-built">How it was built / 制作方式</h2>Before collecting the final dataset, a detailed screening plan was created specifically around the motion problems targeted by this project. 在正式筛选素材前，先根据这个项目想解决的动态问题，专门制定了一套详细筛选方案。Every candidate clip was manually reviewed frame by frame, especially around motion / interaction / contact moments, and unsuitable material was removed one by one. 所有候选片段都会由我人工逐帧观看，重点检查动作 / 交互 / 接触过程，并逐条剔除不符合要求的素材。The remaining clips were then reviewed again with large-model AI under strict project-specific rules, followed by a second filtering pass. 剩余素材还会再由大模型 AI 按照本项目专门制定的严格规则进行第二轮复查与筛选。Only 100 clips were ultimately retained, followed by a project-specific captioning strategy and final training. 最终只有 100 段片段被保留下来，再经过专门为本项目定制的特化打标策略进行标注与训练。<hr /><h2 id="important">Important / 重要说明</h2>This LoRA is not a universal “+20% motion quality” switch. If H3 already generates the scene correctly, you may barely notice it. Its value becomes much easier to see when the original result contains a brief failure. 它并不是一个“打开以后所有动态统一 +20%”的开关。如果 H3 原本就生成得很好，你可能几乎感觉不到它；真正容易看出区别的，是原版已经出现短暂异常的场景。Very high weights are experimental / they may reshape the overall action logic / introduce stronger CG or 3D-game characteristics in extreme dynamic scenes / or affect generated audio. 高权重属于实验区 / 可能重构整体动作逻辑 / 在极端动态场景中增加 CG 或 3D 游戏感 / 也可能影响生成声音。<blockquote>Don't let one bad second ruin an otherwise great clip. 别让不到 1 秒的掉链子，毁掉整段本来很好的视频。</blockquote>

BUNNY General Motion Continuity Repair

Motion_Repair.safetensors

Example Workflow.json

v1.0

🚀 Run It on RunningHub🎁 New users on RunningHub International can get 1000 RH Coins after registration.<a target="_blank" rel="ugc" href="https://www.runninghub.ai/zh-cn/post/2095927034992533505/?inviteCode=lk4qa7rh">🎬 NEW — BUNNY H3 High-Dynamic Two-Pass Latent Upscale Workflow</a> <a target="_blank" rel="ugc" href="https://www.runninghub.ai/zh-cn/post/2095927034992533505/?inviteCode=lk4qa7rh">🎬 新版——BUNNY H3高动态双采潜空间放大工作流</a> It reduces the noise, blur and unstable details that may appear after secondary latent upscaling. 它主要解决动作LoRA在二次放大后容易出现的噪点、模糊和细节不稳定问题。 It supports direct output at up to 1080P and has been optimized for 12GB VRAM. 工作流最高支持1080P直出，并且针对12GB显存进行了优化。 For 12GB VRAM, the recommended setting is 0.3MP for Stage 1 with 1.5× latent upscaling. 12GB显存推荐使用一采0.3MP搭配1.5倍latent放大。<a target="_blank" rel="ugc" href="https://www.runninghub.ai/zh-cn/post/2093321887345819650/?inviteCode=lk4qa7rh ">▶ Start / End Frame One-Click Workflow</a><a target="_blank" rel="ugc" href="https://www.runninghub.ai/zh-cn/post/2093328847935950849/?inviteCode=lk4qa7rh ">▶ Multi-Reference One-Click Workflow</a><a target="_blank" rel="ugc" href="https://www.runninghub.ai/zh-cn/user-center/2092843565587107841/model?inviteCode=lk4qa7rh">📦Bunny Official RunningHub Model Page</a>Future FourBunny LoRAs will also be released and updated on RunningHub.<hr /><h3 id="highly-recommended-for-complex-action:-pair-this-lora-with-bunny-h3-conditioning-bridge-for-better-multi-character-logic-action-ownership-and-scene-stability.-bunny-h3-conditioning-bridge"><a rel="ugc" href="https://civarchive.com/models/2936635/bunny-h3-conditioning-bridge-or-multi-character-logic-and-scene-stability">Highly recommended for complex action: pair this LoRA with BUNNY H3 Conditioning Bridge for better multi-character logic, action ownership, and scene stability. 复杂动作场景强烈推荐搭配 BUNNY H3 Conditioning Bridge，可进一步改善多人互动逻辑、动作归属和场景稳定性。</a></h3><hr /><h2 id="combat-base-v2-is-a-major-upgrade-over-v1-focusing-not-only-on-faster-combat-but-on-better-action-continuity-hit-reactions-cumulative-damage-spatial-movement-and-physical-cause-and-effect.">Combat Base V2 is a major upgrade over V1, focusing not only on faster combat, but on better action continuity, hit reactions, cumulative damage, spatial movement, and physical cause-and-effect.</h2><h3 id="combat-base-v2-is-no-longer-just-a-combat-lora-it-now-works-for-both-action-and-dialogue-scenes-adding-richer-body-movement-finer-sound-details-and-stronger-shot-to-shot-continuity-to-character-performance.combat-base-v2-lora">Combat Base V2 is no longer just a combat LoRA — it now works for both action and dialogue scenes, adding richer body movement, finer sound details, and stronger shot-to-shot continuity to character performance. Combat Base V2 不再只是一个武打 LoRA——它现在可文可武，在文戏中也能增加更丰富的身体动作、更细致的声音表现，以及更强的镜头与状态连续性。</h3><hr /><h3 id="minimax-h312gb-vramuniversal-fl2varef2va-base-model-+-low-sigma-combat-second-pass-workflow"><a target="_blank" rel="ugc" href="https://civarchive.com/models/2869434/minimax-h312gb-vramuniversal-fl2varef2va-base-model-low-sigma-combat-second-pass-workflow?modelVersionId=3241900">【MiniMax H3】12GB VRAM｜Universal FL2VA/REF2VA Base Model + Low-Sigma Combat Second-Pass Workflow</a></h3><h2 id="core-physics-and-motion-logic">Core Physics &amp; Motion Logic / 核心物理与动作逻辑</h2><ul><li>Combat Base V2 moves away from "attack → reset" loops into a continuous physical chain where missed strikes carry momentum, blocks change direction smoothly, and limbs separate cleanly without sticking.</li><li>Combat Base V2 摆脱了“攻击 → 复位”的回合制循环，形成持续的运动链，招式落空可延续动量，格挡可自然顺势变向，且接触点脱离干净不黏连。</li><li>Hit reactions now persist and accumulate—sustained torso bending, lowered guards, and impaired mobility directly build toward convincing heavy finishers.</li><li>受击状态具备持续性与累计效应，躯干弯曲、护架变慢及移动受限等伤害会自然保留，让最后的重击终结更加合理成立。</li></ul><h2 id="combat-skillset-and-staging">Combat Skillset &amp; Staging / 动作技法与镜头调度</h2><ul><li>Upgrades include continuous rotational kicks, sustained grappling and takedowns, dynamic 3-stage environmental collisions, and multi-phase fall trajectories.</li><li>全面升级连续旋转腿法、持续近身缠抱摔投、三阶段环境碰撞（人体→环境→材质），以及拥有完整轨迹的特技跌落表现。</li><li>Supports cinematic 1vMany choreography via staggered attacker entrances, spatial lane control, continuous protagonist-centered tracking, and motivated cross-cut motion continuity.</li><li>支持电影级“一对多”围攻调度，利用交替切入、环境通道压迫、主角中心跟随镜头以及跨镜动作连贯性提升画面表现。</li></ul><h2 id="dialogue-performance-and-benchmarks">Dialogue Performance &amp; Benchmarks / 文戏表现与测试数据</h2><ul><li>Internal same-seed dialogue tests revealed +19.5% in-shot motion, +34.2% audio RMS, and +22.4% edge sharpness, significantly enhancing physical acting, breathing, and vocal dynamics outside of fight scenes.</li><li>同 Seed 文戏测试显示镜内动作量增加 +19.5%、Audio RMS 提升 +34.2%、边缘锐度提升 +22.4%，大幅增强了武戏之外的肢体表演、呼吸与声音动态。</li></ul><h2 id="workflow-and-trigger-usage">Workflow &amp; Trigger Usage / 工作流与触发词使用</h2><ul><li>For optimal flow, structure prompt timing as ACTION → DIALOGUE; start with No Trigger for maximum anatomical accuracy, add <code>prfight2</code> for higher intensity, or <code>prfight2, prfin1</code> for heavy stunt impact.</li><li>建议将提示词结构设为“动作 → 对白”；优先使用 No Trigger 换取最稳动作与人体，需要强攻击性时加入 <code>prfight2</code>，重终结/特技受击则使用 <code>prfight2, prfin1</code>。</li><li>Includes two specialized combat-director workflows (General &amp; 1vMany) with Chinese instructions that support prompt output in any target language.</li><li>附带“通用”与“一对多”两个专用武戏导演工作流，内置中文说明并支持生成任意语言的最终 Prompt。</li></ul>Direct the fight first. Write the prompt second. 先把战斗导演好，再去写提示词。<hr />Upgrade Notes：测试发现：同一镜头包含打斗和台词时，必须先完整完成击打动作再说台词，否则容易退回底模的慢动作表现。Test finding: When combat and dialogue are in the same shot, complete the attack before the dialogue, or the motion may revert to the base model’s slower behavior.<hr /><h2 id="about-this-lora-lora">About This LoRA｜关于这个 LoRA</h2>先说明一下：这不是一个魔法 LoRA，也不会自动帮你编排完整的武打动作。 First of all: this is not a magic LoRA, and it will not automatically choreograph a complete fight for you.如果你只写 <code>two people fighting</code>、<code>a woman fights a man</code> 之类非常简单的 Prompt，模型很可能只是让两个人高速乱挥拳，最后变成一段很快但没有逻辑的“王八拳”。 You still need to tell the model who attacks first, what they do, where the strike lands, how the opponent dodges or reacts, and why the next attack can naturally follow.Prompt 负责决定“他们怎么打”，LoRA 负责增强“这些动作怎么表现”。 The prompt decides what the fighters do, while the LoRA enhances how those actions are performed.这个 Combat LoRA 前后训练和测试了好几个版本，也尝试过不同的数据比例、战斗速度、受击反应以及终结动作训练方向。 This Combat LoRA went through several training and testing versions with different dataset balances, combat-speed behavior, hit reactions, and finisher-focused training.有些旧版本确实能让战斗明显变快，但高权重下也容易破坏 H3 本身不错的动作流畅度，让视频出现倍速感、模糊或者没有恢复过程的连续乱打。 Some earlier versions could make fights noticeably faster, but at higher weights they could also damage H3's native fluidity and produce sped-up motion, blur, or endless attacks without proper recovery.目前发布的这一版，是我实际测试下来相对最平衡、也比较好用的一版。 The version released here is currently one of the most balanced and practical versions from my testing.它不再单纯追求“打得更快”，而是重点强化 连续战斗、命中反馈、身体失衡、后仰、踉跄、击倒、终结逻辑以及攻击产生的实际结果。 Instead of simply making fights faster, it focuses more on continuous combat, impact reactions, loss of balance, recoil, staggering, knockdowns, finishing logic, and clearer physical consequences after attacks.我的目标是尽量保留 H3 本身已经不错的动作连续性，同时让每一次有效攻击看起来更有重量、更有反馈，也更有“打中了以后真的发生了什么”的感觉。 The goal is to preserve H3's existing motion continuity while making successful attacks feel heavier, more reactive, and more consequential.<hr /><h2 id="trigger-words">Trigger Words｜触发词</h2><h3 id="prfight1-general-combat"><code>prfight1</code> — General Combat｜普通战斗</h3>普通战斗只需要使用 <code>prfight1</code>。 For normal combat scenes, simply use <code>prfight1</code>.<h3 id="prfight1-prfin1-finisher-knockout"><code>prfight1, prfin1</code> — Finisher / Knockout｜终结 / 击倒</h3>当这一段战斗需要一个明确的终结、KO、击倒或决定性收尾时，在 <code>prfight1</code> 后加入 <code>prfin1</code>。 When the fight needs a clear finisher, knockout, takedown, or decisive ending, add <code>prfin1</code> after <code>prfight1</code>.<h3 id="prfight1-prfin1-prslow1-slow-motion-finisher"><code>prfight1, prfin1, prslow1</code> — Slow-Motion Finisher｜慢镜终结</h3>需要在最后一击加入短暂的慢动作强调时，再加入 <code>prslow1</code>。 Add <code>prslow1</code> when you want a brief slow-motion emphasis on the decisive finishing strike.注意：只加入 <code>prslow1</code> 并不代表模型一定会自动产生慢动作。 Important: simply adding <code>prslow1</code> does not guarantee that slow motion will automatically appear.你还需要在 Prompt 里明确告诉模型：最后一击在命中瞬间进入短暂慢动作，然后恢复正常速度。 You also need to explicitly tell the model that the decisive final strike receives a brief slow-motion emphasis at the moment of impact, then normal speed resumes.<hr /><h2 id="an-easy-way-to-write-better-h3-fight-prompts">An Easy Way to Write Better H3 Fight Prompts｜最简单的高质量提示词写法</h2>如果你不想自己手动写复杂的动作编排，一个非常简单的方法是直接使用 MiniMax H3 官方的两个 Prompt Writing Guide 文件。 If you do not want to manually write complex choreography, an easy method is to use the two official MiniMax H3 Prompt Writing Guide files.把这两个官方提示词模板文件一起上传给 Gemini。 Upload both official prompt-writing guide files to Gemini.然后告诉 Gemini：按照这两个官方文件的规则，用六段式结构为 H3 编写视频提示词，并告诉它视频长度、剧情和你想要的战斗内容。 Then tell Gemini: follow the rules from these two official files and write an H3 video prompt using a six-part structure, while specifying the video duration, story, and desired fight choreography.通常你只需要告诉 Gemini：几秒、什么剧情、谁打谁、最后想怎么结束，它就可以按照官方模板帮你扩展成一段完整得多的 H3 武打提示词。 Usually, you only need to tell Gemini the duration, basic story, who fights whom, and how you want the sequence to end, and it can expand that into a much more complete H3 fight prompt using the official structure.这也是我目前比较推荐的使用方式：不要让模型自己猜怎么打，而是先让 Gemini 帮你完成动作编排，再把结果交给 H3 + Combat LoRA。 This is also my recommended workflow: do not make the video model guess the choreography; let Gemini build the action sequence first, then send the resulting prompt to H3 + Combat LoRA.<hr /><h2 id="compatibility">Compatibility｜兼容性说明</h2>目前我只在 MiniMax H3 FL2VA 上完整测试过这个 LoRA。 At the moment, I have only fully tested this LoRA with MiniMax H3 FL2VA.其他 H3 模型、不同生成模式或其他工作流我暂时没有进行系统性测试。 I have not systematically tested it with other H3 variants, generation modes, or workflows.所以在非 FL2VA 环境中能否获得相同的效果，我目前无法保证。 Therefore, I cannot currently guarantee identical behavior outside the FL2VA setup.

MiniMax H3 Combat BASE ｜Fight Motion / Impact / Drama Booster

Director of “One Against Many”.txt

H3_Combat_V2.safetensors

Action Director.txt

V2 BASE

━━━━━━━━━━━━━━━━━━🐇 BUNNY H3 Conditioning Bridge━━━━━━━━━━━━━━━━━━Built for high-dynamic MiniMax H3 scenes.专门为 MiniMax H3 高动态复杂场景设计。BUNNY H3 Conditioning Bridge V1 focuses on two core issues:V1 核心解决两个问题：🔥 Multi-Character Interaction Logic / 多人互动逻辑🧩 Scene Stability / 场景稳定性When multiple characters move, attack, cross, or interact simultaneously, H3 often gets confused.当多角色同时移动、攻击、换位或互动时，H3 极易混淆逻辑。 BUNNY H3 Conditioning Bridge helps H3 keep track of who is doing what, to whom, and how the scene evolves. BUNNY H3 Conditioning Bridge 帮助 H3 准确锁定谁在对谁做什么，以及空间后续如何延续。━━━━━━━━━━━━━━━━━━🎯 WHAT IT FIXES🎯 核心修复━━━━━━━━━━━━━━━━━━• Action assigned to the wrong character.• 动作挂错角色。• Attacker and target relationships reversed.• 攻防目标倒置。• Weapons or objects changing owners unexpectedly.• 武器或道具突然换主。• Character states breaking after crossing or occlusion.• 换位或遮挡后角色状态崩坏。• Spatial continuity breaking during fast motion.• 高速位移时空间连续性断层。• Background distortion or unrelated artifacting.• 背景无故畸变或生成无关杂质。━━━━━━━━━━━━━━━━━━⚔️ BEST USE CASES⚔️ 推荐场景━━━━━━━━━━━━━━━━━━✔ 1 vs N &amp; multi-character combat / 1 对多与多人战斗✔ Rapid position swapping &amp; occlusion / 快速换位与遮挡重现✔ Weapon interactions &amp; object ownership / 武器抢夺与物体归属✔ Complex spatial transitions &amp; long action chains / 复杂空间变换与长动作链✔ Multiple characters with distinct visual traits / 多不同服装特征的角色同框Note: Simple single-character scenes will show minimal difference.注：简单单人构图增益有限，场景逻辑越复杂，效果越明显。━━━━━━━━━━━━━━━━━━🎬 DEMO COMPARISONS🎬 实测对比━━━━━━━━━━━━━━━━━━🔧 SEMANTIC BRIDGE vs MOTION CONTINUITY REPAIR━━━━━━━━━━━━━━━━━━They fix different layers of problems. They work great together!它们解决不同维度的痛点，推荐搭配使用！⚡ Motion Continuity Repair→ Smoother action &amp; clearer high-speed motion→ 动作更流畅，高频衔接更完整（负责“怎么动得更好”）🧠 BUNNY H3 Conditioning Bridge V1→ Precise action targets, spatial relationships, and scene stability→ 锁定动作归属、攻防关系与场景逻辑（负责“谁在对谁做什么”）━━━━━━━━━━━━━━━━━━📊 TEST RESULTS &amp; STRENGTH📊 实测数据与推荐参数━━━━━━━━━━━━━━━━━━🟢 60% — Clear logic repair or improvement / 明显修复或增益⚪ 20% — No obvious change / 无显著变化🔴 10% — Minor regressions or new artifacts / 可能引入新异常🎚️ Recommended Strength / 推荐权重:<code>0.10 – 0.15</code> (Start at 0.10. Higher is not always better.)（推荐起始值 0.10，按需微调，并非越高越好）━━━━━━━━━━━━━━━━━━📦 INSTALLATION📦 快速安装━━━━━━━━━━━━━━━━━━1️⃣ Search and install via ComfyUI Manager: <code>BUNNY H3 Semantic Bridge</code>1️⃣ 打开 ComfyUI Manager 搜索并安装：<code>BUNNY H3 Semantic Bridge</code>2️⃣ Download weight file: <code>BUNNY_H3_ActionLogic_Bridge_V1.safetensors</code>2️⃣ 下载权重文件：<code>BUNNY_H3_ActionLogic_Bridge_V1.safetensors</code>3️⃣ Place model in: <code>ComfyUI/custom_nodes/</code>BUNNY_H3_Conditioning_Bridge<code>/models/</code>3️⃣ 放置路径：<code>ComfyUI/custom_nodes/</code>BUNNY_H3_Conditioning_Bridge<code>/models/</code>4️⃣ Drag the example workflow into ComfyUI to start.4️⃣ 直接把页面附加文件中的工作流拖入 ComfyUI 即可使用。━━━━━━━━━━━━━━━━━━ 🖥️ TRAINING INFORMATION 🖥️ 训练信息 ━━━━━━━━━━━━━━━━━━This V1 research was mainly completed on an NVIDIA RTX 6000D 84GB. 这个 V1 的研究与训练主要使用 NVIDIA RTX 6000D 84GB 完成。One Action Logic screening stage evaluated approximately 1,440 prompts across multiple SenseNova → H3 layer combinations. 其中一次 Action Logic 筛选阶段大约测试了 1,440 条 Prompt，并比较了多组 SenseNova → H3 表示层组合。SenseNova U1.5 was used as the external semantic teacher during V1 research. V1 研究阶段使用 SenseNova U1.5 作为外部语义 Teacher。━━━━━━━━━━━━━━━━━━ 🔗 LINKS &amp; CREDITS 🔗 链接与致谢 ━━━━━━━━━━━━━━━━━━🤗 Hugging Face: <a target="_blank" rel="ugc" href="https://huggingface.co/JOKER141/BUNNY_H3_Conditioning_Bridge">https://huggingface.co/JOKER141/BUNNY_H3_Conditioning_Bridge</a>🐙 GitHub: <a target="_blank" rel="ugc" href="https://github.com/aa335615543-ux/BUNNY_H3_Conditioning_Bridge">https://github.com/aa335615543-ux/BUNNY_H3_Conditioning_Bridge</a>🔬 Technical Note: Independently retrained and packaged. Does not include original released weights. 本版本基于公开路线独立重新训练与封装，不包含原作者权重。━━━━━━━━━━━━━━━━━━ 🙏 SPECIAL THANKS 🙏 特别感谢 ━━━━━━━━━━━━━━━━━━Special thanks to speach1sdef178 for publicly sharing the original MiniMax H3 Semantic Bridge research direction, code, data and experiments. 特别感谢 speach1sdef178 公开 MiniMax H3 Semantic Bridge 的原始研究方向、代码、数据和实验过程。🎨 Original Author — Civitai 🎨 原作者 Civitai<a target="_blank" rel="ugc" href="https://civarchive.com/models/2915331/minimax-h3-semantic-bridge?modelVersionId=3298118">https://civarchive.com/models/2915331/minimax-h3-semantic-bridge?modelVersionId=3298118</a>🤗 Original Project — Hugging Face 🤗 原项目 Hugging Face<a target="_blank" rel="ugc" href="https://huggingface.co/speach1sdef178/MiniMax-H3-Semantic-Bridge">https://huggingface.co/speach1sdef178/MiniMax-H3-Semantic-Bridge</a>━━━━━━━━━━━━━━━━━━🚀 NEXT STEP: V2 ROADMAP🚀 下一步：V2 路线图━━━━━━━━━━━━━━━━━━We are exploring dual training tracks for V2 to find the optimal solution:V2 将采用“双轨并行”路线进行对比测试，寻找最优解：Track A — Native H3 Self-Training / 轨 A：H3 原生 Qwen 自训练Explore self-training directly using H3's native text encoder to reduce external dependency.直接基于 H3 原生文本编码器进行自训练，探索降低外部 Teacher 依赖的可能性。Track B — Advanced External Teacher Distillation / 轨 B：进阶外部 Teacher 蒸馏Continue leveraging stronger external semantic models for representation alignment.继续引入更强能力的外部语义 Teacher 模型进行表征对齐与蒸馏。🔥 ~160,000 High-Dynamic Semantic Prompts planned for both tracks.🔥 计划为两条路线准备约 160,000 条高动态语义数据集进行横向对比。Covering combat, parkour, collisions, spatial logic, and state inheritance.覆盖战斗、跑酷、碰撞、空间逻辑、状态继承等全维复杂场景。Goal: Compare both approaches to see which delivers better semantic stability.终极目标：严格对比测试两种路径，找出能让 H3 真正搞懂“谁、在何处、对谁、做了什么”的最优方案。🐇 BUNNY

BUNNY H3 Conditioning Bridge | Multi-Character Logic & Scene Stability

bunnyH3ConditioningBridge_v10.safetensors

bunnyH3ConditioningBridge_v10.json

Files with SHA256: 7042358af063cc25664d5235bab1d2632eb52a9fb3f0a948d65b97091a8688d7