权重简介 / Weights (v2 recommended)

v2(推荐) 于 2026-09-25 在云端用真实 MiniMax-H3 文本编码器(Qwen3-VL-32B int8_convrot,embedding dim=5120)对逻辑链 + XYZ 训练对(wushu_pairs_v2_logic_chains.jsonl,615 pairs,DEFAULT_OPS 含 xyz_drift 与六大翻车)编码后训练:

  • wushu_bridge_wushu_v2.safetensors — residual semantic bridge(trans 2L d=256)
  • wushu_jev_wushu_v2.safetensors — JEV head(val_auc0.997 / val_acc0.984 / ECE~0.001)

v1 文件保留作对照;请优先加载 v2。同路径也在 plugin/ComfyUI-H3-WushuBridge/models/wushu_bridge/。

ComfyUI-H3-WushuBridge · MiniMax H3 武打语义逻辑桥

v1.1.2(XYZ 坐标锁定 + 逻辑链 + 六大翻车差向量)
Hugging Face 仓库现已托管 完整插件 + 权重(原名 h3-wushu-bridge-weights 已重定向到本仓)。
代码功能已到 v1.1.x;已发布的神经网络权重仍为 v1,需本机用 H3 CLIP(5120-d)按 docs/训练流程.md 重训后才能吃满逻辑链 / XYZ。

中文 | English

GitHub(源码) https://github.com/Jojocodex-dotcom/ComfyUI-H3-WushuBridge
本仓(插件 + 权重) https://huggingface.co/Jojocodex/ComfyUI-H3-WushuBridge
配套武打 LoRA https://huggingface.co/Jojocodex/wushu-action-v7-minimax-h3-fl2va-ref2va-lora
许可 Apache-2.0(插件代码与本仓权重;上游 MiniMax H3 另受其社区许可约束)

插件简介

它是什么

ComfyUI-H3-WushuBridge 是插在 MiniMax H3 CONDITIONING 节点之间的残差「武打语义逻辑桥」:

CLIPTextEncode(H3) ──► [ H3 武打语义逻辑桥 ] ──► KSampler / H3 采样节点
        │                        ▲
        │                        └─ 只替换 embedding 张量
        └─ conditioning                 minimax_keyframes / minimax_refs /
           [1, T, 5120] + metadata       minimax_frame_count 等 metadata 原样保留

它把「力线、格距、面对面、命中反馈、切镜身份/空间连续、XYZ 站位」等武打逻辑写进 H3 自己的 conditioning 空间,而不是改提示词字面。架构对标社区 MiniMax-H3-Semantic-Bridge,但训练数据、判据、评分头全部换成武打专用,模型极小,可在装了 H3 的 ComfyUI 里原地训练。

为什么需要它

用户实测中,裸 H3(即便叠武打 LoRA)常出现这些逻辑翻车:

  1. 无意义跳跃 — 腾空无起落力
  2. 未面对面 — 对打却侧身/背对
  3. 法术打空气 — 施法方向不对准对手
  4. 切镜换人 / 位置瞬变 — 身份与站位跳变
  5. 空间逻辑弱 — 左右、间距、朝向混乱(v1.1.2 用显式 xyz= 锁定)
  6. 法术击中无反馈 — 命中后无踉跄/灼衣/击退等物理反应

本插件用「正例种子 − 定向降级负例」把上述失败写成训练差向量;评分头也会对正例扣分。另有 BUNNY 风格高动态逻辑链(缴械、撞墙反弹、遮挡再识别、追击、状态继承等)。

架构要点

项 说明
默认桥 Transformer 残差桥:2 层、d=256、4 头(跨 token,约 4.2M / fp32 ~16MB)
可选桥 MLP 5120→512→512→5120(社区同构,可加载社区权重对照)
应用 out = x + α · residual(x),支持幅度对齐、auto_alpha(分低多修)、guard(改坏回退)
元数据 只动 embedding;minimax_* 等 conditioning metadata 原样保留
评分头 JEV 式:conditioning → P(武打逻辑合格),无解码

节点一览

日常出片默认只注册 5 个核心节点(分类 MiniMax H3/Wushu Bridge):

节点 用途
H3 武打语义逻辑桥 夹在 conditioning 之间,写入武打逻辑残差
H3 武打提示词体检 h3lint 八维规则体检(纯 CPU)
H3 武打逻辑评分 JEV 式概率打分 / 门控
H3 武打编排 动作导演编排
H3 武打桥 清空缓存 卸载常驻权重

训练 / 诊断节点(构建数据集、采集训练对、训桥、训 JEV、降级预览、token 段定位等)默认隐藏。需要时:

# Linux / macOS
export WUSHU_BRIDGE_NODES=all
# Windows cmd
set WUSHU_BRIDGE_NODES=all

然后重启 ComfyUI。

逻辑链 · 降级 · CRITICAL

  • 逻辑链(logic_chains.py):打斗完整弧(逼近→攻防→接触→反馈→终结)+ 行为连续性(ownership / occlusion / momentum / pursuit / facing / state_carry …)+ xyz_coord_duel
  • 降级算子:经典逻辑 ∪ 高动态 ∪ CRITICAL(facing_break / jump_orphan / spell_miss_target / identity_drift / teleport_cut / spell_no_feedback / **xyz_drift**)
  • 默认档案 DEFAULT_OPS 已混入上述 CRITICAL;详见仓库 docs/逻辑链说明.md

XYZ 坐标约定(v1.1.2)

相机相对地面坐标系(建议每镜开头重申):

轴 含义 符号
X 画面左右 左 (−) / 右 (+)
Y 纵深 近相机 (−) / 远 (+)
Z 离地高度 0 = 站立双脚着地

单位:抽象「步」。典型对决:角色A@xyz=(-2,0,0) 与 角色B@xyz=(2,0,0) 面对面。
降级 xyz_drift 会扰动或剥掉坐标;评分 xyz-lock 偏好带 xyz 的多镜对决。

安装路径(全部使用本仓新 URL)

任选其一,放到 ComfyUI/custom_nodes/ComfyUI-H3-WushuBridge/ 后重启:

  1. 插件目录(本仓)
    https://huggingface.co/Jojocodex/ComfyUI-H3-WushuBridge/tree/main/plugin/ComfyUI-H3-WushuBridge
  2. 发行 zip
    https://huggingface.co/Jojocodex/ComfyUI-H3-WushuBridge/resolve/main/releases/ComfyUI-H3-WushuBridge-plugin.zip
  3. GitHub 克隆
    git clone https://github.com/Jojocodex-dotcom/ComfyUI-H3-WushuBridge.git

权重放到 ComfyUI/models/wushu_bridge/(也可直接用插件树内 models/wushu_bridge/ 镜像)。

自检(无需 H3 / 显卡):

python ComfyUI/custom_nodes/ComfyUI-H3-WushuBridge/tools/selftest.py

仓库布局(本 HF 仓)

ComfyUI-H3-WushuBridge/          ← 本 Hugging Face 仓根
├─ README.md                     ← 本模型卡
├─ wushu_bridge_wushu_v1.safetensors
├─ wushu_jev_wushu_v1.safetensors
├─ *_report.json
├─ datasets/
│  └─ wushu_pairs_v2_logic_chains.jsonl   ← 最新 TEXT 对(重训用)
├─ plugin/ComfyUI-H3-WushuBridge/         ← 完整插件源码树
├─ releases/ComfyUI-H3-WushuBridge-plugin.zip
└─ laya/                                  ← 可选 Laya 裁判 bundle

GitHub 源码树布局见项目 README「目录结构」一节。


权重简介

重要:下列 v1 safetensors 在完整逻辑链 / XYZ / 六大翻车差向量落地之前训成。
代码已是 v1.1.2,但神经网络权重尚未用 wushu_pairs_v2_logic_chains.jsonl + H3 CLIP 5120-d 重训。
要吃满新逻辑,请在本机按 plugin/.../docs/训练流程.md 重训后替换。

wushu_bridge_wushu_v1.safetensors(~16 MB)

用途 残差语义桥权重(武打 v1):插在 conditioning 上,把「坏逻辑」embedding 朝「好逻辑」方向推
架构 默认 trans:2 层 Transformer,d=256,4 头;dim=5120(H3)
训练快照 ~4.2M 参数;676 TEXT 对时代训成;验证相对增益 / 方向对齐见同目录 wushu_bridge_wushu_v1_report.json
如何加载 节点「H3 武打语义逻辑桥」→ bridge 选此文件;建议 alpha≈0.12、magnitude_match=per_token、token_span=all(参考图模式用 tail)
行为开关 alpha 固定强度;auto_alpha 按 JEV 分偏低时加大修正;guard 若改坏评分则自动回退
尚不包含 逻辑链 / CRITICAL(含 xyz_drift)全量重训结果;不是完整 H3 或 LoRA

wushu_jev_wushu_v1.safetensors(~6.7 MB)

用途 JEV 式评分头:conditioning → P(武打逻辑合格),供评分节点与桥的门控 / auto_alpha
架构 hidden=256,约 1.7M 参数;温度校准后 ECE≈0.009;验证准确率≈0.83、AUC≈0.92(见 report)
如何加载 「H3 武打逻辑评分」→ judge 选此文件,aggregate=mean,threshold=0.5;桥节点也可同时挂上作 guard
尚不包含 针对 v1.1+ 六大翻车 / XYZ 槽位的重标定

*_report.json

训练指标快照(参数量、验证 sem / accuracy / AUC / ECE、漂移、逐轮 history)。便于对照复现,不是推理必需。

datasets/wushu_pairs_v2_logic_chains.jsonl(~615 对)

用途 最新 TEXT 好坏对,带逻辑链标注与 CRITICAL 降级(含 xyz_drift)
用途边界 供重训桥与 JEV;不是给推理节点直接加载的权重
配套 datasets/wushu_pairs_v2_logic_chains_stats.json

laya/(若存在)

可选 Laya 裁判 bundle(english + multilingual)。装完插件后:

python tools/setup_laya.py --from ours

即可从本仓 laya/ 装配,无需手抠文件。

已移除(obsolete)

以下文件已从本仓删除,请勿再引用旧链接:

  • wushu_bridge_cloud.safetensors / wushu_jev_cloud.safetensors(早期云端对照)
  • wushu_pairs_v1_pairs.jsonl / wushu_pairs_v1.json(旧对清单;请改用 v2 logic_chains)

旧仓名 Jojocodex/h3-wushu-bridge-weights 已重定向到 **Jojocodex/ComfyUI-H3-WushuBridge**。


快速开始

# 1) 装插件(zip 示例)
cd ComfyUI/custom_nodes
curl -L -o plugin.zip \
  https://huggingface.co/Jojocodex/ComfyUI-H3-WushuBridge/resolve/main/releases/ComfyUI-H3-WushuBridge-plugin.zip
unzip plugin.zip   # 得到 ComfyUI-H3-WushuBridge/

# 2) 装权重(若 zip/插件树里尚未带上)
mkdir -p ComfyUI/models/wushu_bridge && cd ComfyUI/models/wushu_bridge
BASE=https://huggingface.co/Jojocodex/ComfyUI-H3-WushuBridge/resolve/main
curl -L -O $BASE/wushu_bridge_wushu_v1.safetensors
curl -L -O $BASE/wushu_jev_wushu_v1.safetensors

# 3) 重启 ComfyUI → 接线
# CLIPTextEncode(H3) → H3武打语义逻辑桥 → 采样
# bridge = wushu_bridge_wushu_v1.safetensors
# judge  = wushu_jev_wushu_v1.safetensors(可选)

提示词里可写 XYZ,例如:角色A@xyz=(-2,0,0) / Fighter A at xyz=(-2,0,0)。


链接

引用 / 许可

  • 本插件与本仓发布的 wushu v1 权重:Apache-2.0
  • 内嵌 Laya 源码副本:Apache-2.0(见 plugin/.../wushu_bridge/vendor/laya/LICENSE)
  • MiniMax H3 及衍生模型:遵循上游 MiniMax H3 Community License
@misc{comfyui-h3-wushubridge,
  title        = {ComfyUI-H3-WushuBridge: MiniMax H3 Wushu Semantic Logic Bridge},
  author       = {Jojocodex},
  year         = {2026},
  howpublished = {\url{https://github.com/Jojocodex-dotcom/ComfyUI-H3-WushuBridge}},
  note         = {Weights: \url{https://huggingface.co/Jojocodex/ComfyUI-H3-WushuBridge}}
}

English

What this is

ComfyUI-H3-WushuBridge is a residual wushu (martial-arts fight) semantic logic bridge for MiniMax H3. It sits between CONDITIONING nodes:

CLIPTextEncode(H3) → [ Wushu Bridge ] → sampler

It only replaces the embedding tensor and preserves minimax metadata (minimax_keyframes, minimax_refs, minimax_frame_count, …). Default architecture is a 2-layer Transformer bridge (d=256); an MLP option matches the community Semantic Bridge for A/B loading.

Why

Bare H3 often breaks fight logic: meaningless jumps, fighters not face-to-face, spells missing the opponent, identity/position jumps on cuts, weak spatial logic, weak spell-hit feedback — plus XYZ coordinate locking in v1.1.2.

Six CRITICAL failure modes are encoded as degrade ops (facing_break, jump_orphan, spell_miss_target, identity_drift, teleport_cut, spell_no_feedback, xyz_drift).

Nodes

Core (default): bridge, h3lint, JEV score, choreography, clear-cache.
Train/dev nodes appear when WUSHU_BRIDGE_NODES=all.

XYZ convention

Camera-relative ground frame: X left(−)/right(+), Y near(−)/far(+), Z height (0 = standing). Example: A@xyz=(-2,0,0) facing B@xyz=(2,0,0).

Install

Weights (detail)

File Size Role
wushu_bridge_wushu_v1.safetensors ~16MB Residual semantic bridge (trans, 2L d=256). Load in bridge node. Supports alpha / auto_alpha / guard.
wushu_jev_wushu_v1.safetensors ~6.7MB JEV-style head: conditioning → P(logic-pass). Score node + gating.
*_report.json small Training metrics snapshots
datasets/wushu_pairs_v2_logic_chains.jsonl ~615 pairs Latest TEXT pairs with logic chains + CRITICAL ops (incl. xyz_drift) — for retraining
laya/ optional Laya judge bundle (tools/setup_laya.py --from ours)

Explicit: v1 weights predate full logic-chain / XYZ retrain. Code is v1.1.x; neural weights need a local retrain with H3 CLIP 5120-d (see docs/训练流程.md).

Removed: cloud weights and v1 pair dumps — use latest files only. Old repo id h3-wushu-bridge-weights redirects here.

License

Apache-2.0 for this plugin and published wushu v1 weights. Upstream MiniMax H3 follows its own community license.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support