ComfyUI-JZL-MiniMax-H3
A ComfyUI extension with 37 custom nodes.
Nodes (37)
The resolution calculator that does H3's grid math so you don't have to
Same H3 calculator, plus the one number the second-pass upscale needs
Split your script into per-shot lines and route them to N outputs
The GGUF loader with every sampling knob, and a backend that doesn't fight ComfyUI
Put your LLM API key in a dialog, not in your workflow JSON
The all-in-one that turns a script, some assets, and H3 into finished clips
Chained clips that don't restart every 5 seconds
Run 48 MiniMax H3 clips from one node, and keep the VRAM bill flat
JZL's director's console, Mini
MiniMax-H3 makes five-second clips. This node turns a paragraph into the whole short.
Route the right reference audio to the right segment, dynamically
The fixed-port audio router — eight named inputs, no surprises
The node that keeps your first/last frames honest after a latent upscale
Encode both passes at once — first-pass latent plus a second-pass conditioning
Tell the script LLM exactly what camera language you want
The scratchpad that inserts H3's six-segment prompt templates for you
Take a ref2va bus apart into the exact ports the encoder wants
Bundle the 18 ref2va ports into one bus that only fits where it should
Unpack one reference bus back into 48 named ports
Squeeze 48 reference assets into one wire
The ref2va encoder that's a 100% replica of ComfyUI's official H3 node — plus a boost
Same ref2va encoder, fixed ports — for workflows that want a stable layout
Turn a stack of music dropdowns into a MiniMax Music3 structured caption
The same Music3 caption builder, but for two singers who need distinct parts
It's a text box with buttons — and that's exactly what Music3 lyrics need
Point-and-click prompt engineering for MiniMax H3 — 39 visual styles, zero API keys
A prompt enhancer that knows exactly what not to touch
The multi-reference H3 preset — where your reference images do the talking
It reads your node names and sorts your reference images for H3
The no-magic scene dispatcher — 24 named ports instead of name-guessing
The one-call brain that turns a story into a full H3 shooting script
Actually read what the LLM wrote, without leaving the canvas
The segment processor that turns an LLM's messy script into four clean lists
Feed it reference videos, and it pairs each one with its audio track
The fixed-port video dispatcher — 8 named videos, each with its audio pre-paired
The 24-port output sorter that hands your H3 clips to Video Combine
A MiniMax H3 screening room on your canvas — no wires attached
ComfyUI-JZL-MiniMax-H3
MiniMax H3 视频生成节点（基于 ComfyUI V3 io.Schema 机制）。
当前节点
| 节点 | node_id | 说明 |
|------|---------|------|
| 🎬 MiniMax H3 参考编码 | JZL_MiniMaxH3ReferenceToVideo | ref2va 参考条件编码，100% 复刻官方 MiniMaxH3ReferenceToVideo |
特性
- V3 原生动态接口：参考图 / 参考视频 / 视频音轨 / 独立音频 四组端口用
io.Autogrow实现「连接后自动长出下一个空槽」，与官方一致。 - 参考值放大（JZL 扩展）：仅
match模式生效，ref_scale范围1.0~5.0，步长0.1。最终像素面积 = 生成画布面积 × 倍率（面积倍率，非分辨率倍率）。1.0= 官方行为。
提示词标签
| 标签 | 含义 |
|------|------|
| <Picture i> | 第 i 张参考图 |
| <Video k> | 第 k 段参考视频 |
| <Audio j> | 第 j 条参考音频（视频音轨会在其 <Video> 前单独成标签） |
安装
把本目录放入 ComfyUI 的 custom_nodes/ 下，或建立符号链接：
New-Item -ItemType Junction -Path "D:\...\ComfyUI\custom_nodes\ComfyUI-JZL-MiniMax-H3" -Target "D:\AI_JZL\ComfyUI-JZL-MiniMax-H3"
⚠️ 解压最常见的坑：不要多套一层目录。 正确的结构是
custom_nodes\ComfyUI-JZL-MiniMax-H3\__init__.py。 若解压后变成custom_nodes\ComfyUI-JZL-MiniMax-H3\ComfyUI-JZL-MiniMax-H3\__init__.py， ComfyUI 找不到入口文件，一个节点都不会出现（日志里是Cannot import ... module for custom nodes）。
加载日志（排查「节点不显示」）
启动时本插件会打印一行加载结果，可直接判断：
- ✅ 成功：
[JZL MiniMax H3] ✅ 加载成功：37 个节点 ｜ v0.9.02 ｜ ComfyUI 0.41.2 ｜ … - ⚠️ 部分失败：
⚠️ 部分模块加载失败（N 组正常…）并在下方逐条列出失败的模块与处置建议（其余节点仍可用） - ❌ 整体失败：
❌ 节点包加载失败：ComfyUI-JZL-MiniMax-H3（节点列表里不会出现本插件）， 附带失败模块 / 原因 / 处置建议 / 排查步骤；核心模块（官方编码链路）失败时该结构才会出现。
一行命令拿到完整环境报告（排障时把这行输出发回即可）：
<你的 ComfyUI python> "custom_nodes\ComfyUI-JZL-MiniMax-H3\jzl_diagnose.py"
依赖
- ComfyUI ≥ 0.29.0（含官方 MiniMax H3 支持；
comfy_api.latest自 0.3.47 起、H3 核心自 0.29.0 起）- 推荐 ≥ 0.37.0：低于它时缺
comfy.audio，音频重采样会自动回退 torchaudio（功能不受影响）。 - ComfyUI 0.41.x 已实测通过（加载 37 个节点）。
- 推荐 ≥ 0.37.0：低于它时缺
torch、torchaudio（缺失或与 torch 版本不匹配会导致本插件整个不加载，日志会直接指出）
本地 LLM 运行时（llama-server）
「模型加载 Pro / 剧本与镜头处理器」的本地 LLM/VLM 推理不再依赖 llama-cpp-python，改用 llama-server 子进程（进程隔离、跑完即停、自动释放显存）。
首次使用本地模型前，请先运行安装器（仅用 Python 标准库，自动识别系统/架构/显卡，下载并校验固定 llama.cpp b10436 预编译运行时，零 pip 依赖）：
- Windows：双击
install_runtime.bat一键安装（自动定位 Python） - 或手动执行：
cd custom_nodes\ComfyUI-JZL-MiniMax-H3
..\..\..\python_embeded\python.exe install_runtime.py # Windows 便携包
# 或：python install_runtime.py
--dry-run只看检测结果与计划，不下载--list-backends查看当前机器可用后端--force强制重装
已安装的
llama-cpp-python无需卸载，也不会冲突（本节点已不再引用它）；想省磁盘可自行pip uninstall llama-cpp-python。