H3 Fight Flow V1 | Continuous Fight Interaction LoRA

Experimental Continuous Fight Interaction LoRA 实验性连续战斗交互增强 LoRA My goal is simple: push H3 from “can fight” to “knows how to fight.” 我的目标很简单：让 H3 从“能打”，变成真正“会打”。 🚀 Run It on RunningHub 🎁 New users on RunningHub International can get 1000 RH Coins after registration. 国内用户可以使用： What is this? 这是干什么的？ Fight Flow V1 does not mainly teach new moves. It explores whether H3 can learn how two or more characters should actually fight together. Fight Flow V1 主要不是教 H3 更多招式，而是在实验：能不能让两个或多个人真正围绕同一场战斗进行互动。 Initiation → response → contact → consequence → initiative shift → next action. 发起 → 回应 → 接触 → 结果 → 主动权转换 → 下一动作继承。 The previous action should create the starting state of the next one, instead of constantly resetting back to a neutral pose. 上一招产生的位置、重心、距离和姿态，应该成为下一招的起点，而不是打一轮就重新摆架势。 D-OPSD Trained D-OPSD 训练版本 Unlike regular LoRA training that mainly relies on captions, D-OPSD allows the Teacher to directly reference the target training video. 不同于主要依赖 Caption 的普通 LoRA，D-OPSD 训练时 Teacher 可以直接参考目标训练视频。 This helps learn motion relationships that are difficult to fully describe with text: contact, trajectory, forced movement, weapon interaction, initiative changes and state transitions. 这样可以学习很多文字难以完整描述的内容，例如接触、轨迹、被动位移、武器交互、主动权变化和动作状态继承。 This V1 was trained entirely on RTX 6000D, with VRAM peaks around 60GB and roughly 2–3× my usual LoRA training time. 这一版全程使用 RTX 6000D 训练，显存峰值约 60GB，整体训练时间大约是我平常 LoRA 的 2～3 倍。 What does it improve? 它主要增强什么？ Better attack-response causality, more natural initiative changes, stronger state continuity, and fewer “attack → reset → attack again” sequences. 攻击与回应的因果更完整，主动权转换更自然，状态继承更明显，也更少出现“打一招 → 重置 → 再打一招”。 In multi-person fights, one opponent can leave the exchange while another enters, instead of everyone simply performing independent actions at the same time. 在多人战中，当前敌人退出、主角换位、下一名敌人接入的衔接也会更加自然，而不是所有人同时各打各的。 Even with only a short prompt + trigger, the LoRA can already organize fairly complete combat exchanges by itself. 目前测试中，即使只使用一句简单 Prompt + 触发词，也已经能够自行组织出比较完整的连续战斗。 It is NOT a speed LoRA 它不是“加速 LoRA” The improvement is not simply “make both characters move faster.” It is mainly about how motion is organized and passed from one action into the next. 它并不是简单把两个人加速，让画面看起来更激烈；真正增强的是动作如何被组织，以及上一动作如何把状态传递给下一动作。 Three interaction triggers 三大交互触发词 - Unarmed combat interaction. — Use when punches, kicks, grappling, throws or body control dominate the fight. Unarmed combat interaction. —— 拳脚、摔投、身体控制等徒手交互主导时使用。 - Melee weapon interaction. — Use when blades, staffs, spears or other melee weapon lines dominate the exchange. Melee weapon interaction. —— 刀、剑、棍、枪等近战武器的攻击线、格挡和兵器接触主导时使用。 - Gun-fu interaction. — Use when muzzle control, firearm handling, close-range shooting or gun-hand interaction dominates the fight. Gun-fu interaction. —— 枪口控制、控枪、近距离射击和枪械近战交互主导时使用。 Choose the trigger by the dominant interaction logic, not simply by what equipment appears on screen. 判断时看整段战斗由什么交互逻辑主导，而不是单纯看角色手里拿了什么。 Body / fists → Unarmed | Melee weapon line → Melee Weapon | Muzzle / firearm control → Gun-fu 身体 / 拳脚主导 → 徒手｜近战武器线主导 → Weapon｜枪口与枪械控制主导 → GunFu Recommended use 推荐使用方式 Fight Flow works as an interaction layer together with Combat, Weapon and Motion Repair LoRAs. Fight Flow 更适合作为交互层，与 Combat、Weapon、GunFu、Motion Repair 等 LoRA 组合使用。 Specialized LoRAs decide what moves the characters know. Fight Flow helps those moves become one continuous fight. 专项 LoRA 决定角色会什么动作，Fight Flow 负责让这些动作真正连接成一场战斗。 Current tested weight: 0.65 in the first pass / 0.25 in the second pass. 当前测试权重：一采 0.65 / 二采 0.25。 Before: two characters are both fighting. 以前：两个人都在打。 Fight Flow V1: they are fighting the same fight. Fight Flow V1：两个人开始真正打同一场架。