BUNNY_H3_Conditioning_Bridge
Semantic Conditioning Bridge for MiniMax H3
Nodes (1)
🐇 BUNNY_H3_Conditioning_Bridge
A semantic-conditioning bridge for MiniMax H3.
一个给 MiniMax H3 使用的语义增强 Bridge。
It is not a LoRA and it does not rewrite your prompt.
它不是 LoRA，也不会改写你的 Prompt。
Its purpose is to help H3 keep complex action relationships clearer.
它的作用是帮助 H3 在复杂动作里更稳定地保持人物、动作、目标、物体和前后状态之间的关系。
✨ What does it improve?
✨ 它主要改进什么？
Complex action can sometimes cause H3 to lose track of who is doing what.
复杂动作里，H3 有时会搞不清谁在做什么。
Typical problems include:
常见问题包括：
-
Actions jumping from one character to another.
动作从一个人串到另一个人。 -
Weapons or objects changing owner.
武器或道具突然换主人。 -
Attacker / target relationships becoming confused.
攻击者和目标关系混乱。 -
Identity relationships breaking after turning, crossing, or occlusion.
转身、换位、遮挡以后人物关系丢失。 -
The next action continuing from the wrong previous state.
后一个动作从错误的前一状态继续。
BUNNY_H3_Conditioning_Bridge is designed to reduce these failures.
BUNNY_H3_Conditioning_Bridge 主要就是针对这些问题。
⚔️ Recommended combinations
⚔️ 推荐搭配
I currently recommend using it together with COMBAT V2 or Motion Continuity Repair LoRA.
目前推荐和 COMBAT V2 或 动作连续性修复 LoRA 一起使用。
COMBAT V2 / Motion Continuity Repair
→ how the action moves and continues
→ 负责动作怎么打、怎么接、怎么保持连续
Semantic Bridge V1
→ who is doing what and which state belongs to whom
→ 负责谁在做什么、动作和状态到底属于谁
This is especially useful in complex multi-character scenes such as 1-vs-4 combat.
在 1 对 4 这种复杂多人场景里尤其有帮助。
It can help keep the scene more stable when several characters move, exchange positions, attack, react, or occlude each other.
多人同时移动、换位、攻击、受击、互相遮挡时，它能帮助 H3 更稳定地保持场景关系。
📦 Installation
📦 安装
1. Custom node
1. 自定义节点
Put:
把：
BUNNY_H3_Conditioning_Bridge
into:
放到：
ComfyUI/custom_nodes/
Then restart ComfyUI.
然后重启 ComfyUI。
2. Model
2. 模型文件
Model:
模型：
BUNNY_H3_ActionLogic_Bridge_V1.safetensors
Recommended folder:
推荐目录：
ComfyUI/custom_nodes/BUNNY_H3_Conditioning_Bridge/models/
If installed through ComfyUI Manager, the folder name may appear as bunny-h3-semantic-bridge. This is normal.
如果通过 ComfyUI Manager 安装，节点文件夹名称可能会显示为 bunny-h3-semantic-bridge，这是正常的。
The node does not depend on a fixed folder name. Just place the Bridge model inside the models folder of the actual installed node directory.
节点并不依赖固定的文件夹名称，只需要把 Bridge 模型放进你实际安装的节点目录中的 models 文件夹即可。
例如：
ComfyUI/custom_nodes/BUNNY_H3_Conditioning_Bridge/models/
或者：
ComfyUI/custom_nodes/bunny-h3-semantic-bridge/models/
Both are valid. 两种都可以。
⚙️ Usage
⚙️ 使用方法
Use the node:
使用节点：
BUNNY H3 Conditioning Bridge
Recommended starting settings:
推荐起始参数：
alpha = 0.10 ~ 0.15
magnitude_match = per_token
enabled = true
Start low and test with the same seed.
建议从低强度开始，并使用同一个 Seed 做对比。
Higher strength is not automatically better.
强度并不是越高越好。
🔌 Where should the node be connected?
🔌 节点应该接在哪里？
The Bridge must be placed after the MiniMax H3 text-conditioning encoder.
Bridge 必须放在 MiniMax H3 文本 Conditioning 编码完成之后。
The easiest way to use it is to find the existing CONDITIONING cable in your H3 workflow, disconnect it, insert the Bridge in the middle, and reconnect the Bridge output to the original downstream node.
最简单的用法就是：找到你 H3 工作流里原本那根 CONDITIONING 线，把它断开，把 Bridge 插在中间，再把 Bridge 输出重新接回原来的下游节点。
BEFORE / 原来：
H3 Text / Prompt Conditioning
│
│ CONDITIONING
▼
Original downstream node
原来的下游节点
AFTER / 加入 Bridge 后：
H3 Text / Prompt Conditioning
│
│ CONDITIONING
▼
BUNNY H3 Conditioning Bridge
│
│ CONDITIONING
▼
Original downstream node
原来的下游节点
In simple terms:
简单来说：
H3 Conditioning
→ BUNNY H3 Conditioning Bridge
→ Original downstream node
H3 Conditioning
→ BUNNY H3 Conditioning Bridge
→ 原来要接的下游节点
Do not connect the Bridge to MODEL, LoRA, LATENT, VAE, image, or audio lines.
不要把 Bridge 接到 MODEL、LoRA、LATENT、VAE、图片或音频线路上。
It only receives and outputs CONDITIONING.
它只接收并输出 CONDITIONING。
If your workflow already works normally without the Bridge, you do not need to rebuild the workflow. Just insert the Bridge into the existing H3 CONDITIONING connection.
如果你的工作流本来就可以正常运行，不需要重新搭工作流，只需要把 Bridge 插进现有的 H3 CONDITIONING 线路中间即可。
📊 Current test observations
📊 当前测试观察
In my current test set, about 60% of cases showed different levels of repair or improvement.
在我目前的测试里，大约 60% 的案例出现了不同程度的修复或增益。
About 20% showed no obvious difference between Bridge OFF and Bridge ON.
大约 20% 的案例在 Bridge OFF / ON 之间没有明显差别。
Around 10% produced new errors or regressions after enabling the Bridge.
大约 10% 的案例在开启 Bridge 后会出现新的错误或退化。
These are approximate observations from my current tests, not a formal benchmark or guaranteed success rate.
这些数字只是目前测试中的大致观察，不是正式 Benchmark，也不代表固定成功率。
Because these are approximate field observations rather than a normalized statistical report, the percentages are not intended to sum exactly to 100%.
因为这是实际测试中的近似观察，而不是严格归一化的统计报告，所以这些比例不要求精确相加等于 100%。
📥 Model links
📥 模型地址
Semantic Bridge V1 and work flow:
Semantic Bridge V1 和工作流：
https://huggingface.co/JOKER141/BUNNY_H3_Conditioning_Bridge
COMBAT V2:
COMBAT V2：
https://huggingface.co/JOKER141/MiniMax-H3-Combat-Base-V2
Motion Continuity Repair LoRA:
动作连续性修复 LoRA：
https://huggingface.co/JOKER141/MiniMax-H3-General-Motion-Continuity-Repair
🙏 Original project
🙏 原项目
This project was originally inspired by speach1sdef178 / MiniMax-H3-Semantic-Bridge.
这个项目最早受 speach1sdef178 / MiniMax-H3-Semantic-Bridge 启发。
Thanks to the original author for publicly sharing the research idea, code, data, and experiments.
感谢原作者公开研究思路、代码、数据和实验过程。
Original project:
原项目：
https://huggingface.co/speach1sdef178/MiniMax-H3-Semantic-Bridge
The LoRA helps H3 perform the motion. Semantic Bridge V1 helps H3 keep track of who is doing what while the scene gets complicated.
LoRA 负责让 H3 把动作做出来，Semantic Bridge V1 负责场景复杂以后，尽量别搞丢谁在干什么。
FourBunny / JOKER141