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

━━━━━━━━━━━━━━━━━━

🐇 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 & multi-character combat / 1 对多与多人战斗

✔ Rapid position swapping & occlusion / 快速换位与遮挡重现

✔ Weapon interactions & object ownership / 武器抢夺与物体归属

✔ Complex spatial transitions & 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 & 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 & 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:

0.10 – 0.15 (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: BUNNY H3 Semantic Bridge

1️⃣ 打开 ComfyUI Manager 搜索并安装：BUNNY H3 Semantic Bridge

2️⃣ Download weight file: BUNNY_H3_ActionLogic_Bridge_V1.safetensors

2️⃣ 下载权重文件：BUNNY_H3_ActionLogic_Bridge_V1.safetensors

3️⃣ Place model in: ComfyUI/custom_nodes/BUNNY_H3_Conditioning_Bridge/models/

3️⃣ 放置路径：ComfyUI/custom_nodes/BUNNY_H3_Conditioning_Bridge/models/

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 & CREDITS 🔗 链接与致谢 ━━━━━━━━━━━━━━━━━━

🤗 Hugging Face: https://huggingface.co/JOKER141/BUNNY_H3_Conditioning_Bridge

🐙 GitHub: https://github.com/aa335615543-ux/BUNNY_H3_Conditioning_Bridge

🔬 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

https://civitai.red/models/2915331/minimax-h3-semantic-bridge?modelVersionId=3298118

🤗 Original Project — Hugging Face
🤗 原项目 Hugging Face

https://huggingface.co/speach1sdef178/MiniMax-H3-Semantic-Bridge

━━━━━━━━━━━━━━━━━━

🚀 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