|
Download README.md from taurusduan/MiniMax-H3-General-Motion-Continuity-Repair: direct link, hf CLI and curl.
- Browser
- Download file 10.8 kB
-
https://huggingface.co/taurusduan/MiniMax-H3-General-Motion-Continuity-Repair/resolve/main/README.md
- Command line
-
hf download hf://taurusduan/MiniMax-H3-General-Motion-Continuity-Repair/README.md
-
curl -L -o README.md https://huggingface.co/taurusduan/MiniMax-H3-General-Motion-Continuity-Repair/resolve/main/README.md
10.8 kB
| language: | |
| - en | |
| - zh | |
| base_model: | |
| - MiniMaxAI/MiniMax-H3 | |
| tags: | |
| - MINIMAX | |
| - MINIMAXH3 | |
| - MOTION | |
| - lora | |
| license: other | |
| license_name: minimax-h3-community-license | |
| license_link: https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE | |
| General Motion Continuity Repair LoRA | |
| 通用动作连续性修复 LoRA | |
| 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。 | |
| *🎬 BUNNY Motion Repair V2 —— 高难动作不该再靠抽卡!* | |
| *BUNNY Motion Repair V2 — High-Difficulty Motion Shouldn't Be a Lottery.* | |
| *🎯 核心升级 | What's New* | |
| *⚡ 专攻“极危禁区”: 重点修复 MiniMax H3 过去最易崩溃的高速空翻、旋转、翻滚、倒置及恢复。大幅增加 Full-Body 倒置旋转训练,同时延续 V1 的泛用修复能力。* | |
| *Fixes H3's hardest failure zones: flips, spins, rolls, inversions, and orientation recovery, with expanded rotational training while keeping V1's general power.* | |
| *🌪️ 地面支撑与空中翻腾双提升: 回旋、风车、Airflare、转体跳跃与落地缓冲稳定性暴涨!告别人体扭成一团、肢体接错或头躯不同步。* | |
| *Major stability gains for flares, windmills, airflares, spins, and landing recovery. Stops limbs reconnecting wrong or bodies twisting into unreadable masses.* | |
| *🔄 全链路衔接优化: 强化起跳 ➔ 倒置 ➔ 连续旋转 ➔ 重新找正 ➔ 落地全过程,不只是中段好看。* | |
| *Improves the full motion sequence from takeoff, inversion, and spinning to reorientation and landing.* | |
| *💡 底层哲学 | Pure Capability* | |
| *🧠 能力导向打标: 学习的是人体结构、朝向感知与动量恢复而非死记具体招式,拥有极强泛化能力。* | |
| *Trains core mechanics—structure, momentum, and recovery—rather than specific trick names for maximum generalization.* | |
| *🎬 从废片到可用: 依然是修复 LoRA(极端过渡帧仍可能偶发小瑕疵),但足矣让输出从“直接废弃”变成“直接可用”。* | |
| *Still a repair tool, but turns discarded clips into usable final shots.* | |
| --- | |
| Recommended Usage / 推荐使用 | |
| Standalone use: around 0.9 is currently recommended. | |
| 单独使用时,目前推荐权重约为 0.9。 | |
| 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。 | |
| With a Combat LoRA: Stage 1 around 0.5–0.7 is recommended. | |
| 搭配 Combat LoRA 时,一采推荐约 0.5–0.7。 | |
| 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 0.2–0.3 is recommended. | |
| 二采建议约 0.2–0.3。 | |
| 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: | |
| 单独使用:≈0.9 / 搭配 Combat:0.5–0.7 / 二采:0.2–0.3 | |
| 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。 | |
| Trigger Word / 触发词 | |
| Trigger word: bunny_crisp_motion — optional. It mainly serves as an explicit identifier for the LoRA and is not required for normal use. | |
| 触发词:bunny_crisp_motion —— 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。 | |
| What does it do? / 它主要做什么? | |
| 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。 | |
| 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。 | |
| Current A/B Results / 当前对比表现 | |
| 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。 | |
| Weapon Interaction / 武器交互 | |
| 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。 | |
| How it was built / 制作方式 | |
| 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。 | |
| Important / 重要说明 | |
| 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。 | |
| 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。 |