Hugging Face's logo

yj0942
/
H3MD-Director

MiniMax H3
comfyui
video
long-video
workflow
Model card Files
xet
 Community

H3MD Director (H3MD 导演台) · v0.4.3 CCW

Free, open-source ComfyUI nodes that render 2–3 minute videos with MiniMax H3 in one run (1080P; 2K for single 15 s clips). 免费开源的 ComfyUI 节点：用 MiniMax H3 一次直出两三分钟的 1080P 长视频。

This is a download mirror of the node pack. Source code, docs and updates: https://github.com/1113800942/comfy-h3md-director 这里是节点包的下载镜像，源码、文档和更新以 GitHub 为准。

Files / 文件

  • ComfyUI-H3MD-Director_v0.4.3.zip — latest node pack + example workflows + Chinese README + prompt skill / 最新节点包 + 示例工作流 + 中文说明 + 提示词 skill
  • plans/H3MD_demo_eve_mv.zip — demo plan pack: our Eve music video "Above the Clouds" (11 shots, 2 min 40 s; pictures + music + all settings) / 示例计划包：Eve 的 MV《Above the Clouds》（11 个镜头 2 分 40 秒，图、音乐、全部设置）
  • README_EN.md — install guide in English / 英文安装说明
  • H3MD导演台_安装说明.txt — install steps in Chinese / 中文安装步骤
  • ComfyUI-H3MD-Director_v0.4.2.zip, ComfyUI-H3MD-Director_v0.4.0.zip, ComfyUI-H3MD-Director_v0.3.6.zip, ComfyUI-H3MD-Director_v0.3.5.zip, ComfyUI-H3MD-Director_v0.3.4.zip, ComfyUI-H3MD-Director_v0.3.3.zip, ComfyUI-H3MD-Director_v0.3.2.zip, ComfyUI-H3MD-Director_v0.3.1.zip, ComfyUI-H3MD-Director_v0.3.0.zip, ComfyUI-H3MD-Director_v0.2.2.zip, ComfyUI-H3MD-Director_v0.1.1.zip — older versions / 旧版

国内下载

把网址里的 huggingface.co 换成 hf-mirror.com 即可高速下载，不用登录。

Install / 安装

Unzip into ComfyUI/custom_nodes/ (or install "H3MD Director" from ComfyUI Manager / Comfy Registry; check the version shown in the Director node title), ComfyUI 0.43+ (0.44+ for Seamless long video), restart, then drag a workflow from example_workflows/ into ComfyUI – start with H3MD Director simple.json (v0.3: Control Panel + Director + preview). GPU: 16 GB VRAM minimum, 24 GB+ recommended. 解压到 ComfyUI/custom_nodes/（也可以在 ComfyUI 管理器里搜 "H3MD Director" 安装；装好后导演台节点标题会显示版本号），ComfyUI 0.43 或更新（无接缝长视频需要 0.44 或更新），重启后把 H3MD Director simple.json（清爽版） 拖进 ComfyUI。显存最低 16 GB，推荐 24 GB 以上。 Models and low-VRAM settings: see README.md / README.zh-CN.md inside the zip.

First run in one click / 一键跑通

Open H3MD Director simple.json → on the Director Import plan… → pick H3MD_demo_eve_mv.zip → Plan + all settings. Press Run: it renders the first segment only (~8 min on an RTX 5090); press Clear on the orange banner to render the whole video. Missing models: ComfyUI offers the download. 打开 H3MD Director simple.json → 导演台点 Import plan… → 选 H3MD_demo_eve_mv.zip → Plan + all settings。直接点运行：默认只跑第一段（RTX 5090 约 8 分钟）；点总控台橙色提示里的清空就跑整首。缺模型时 ComfyUI 会弹出下载。

What's new / 更新

  • 0.4.3: lighter Control Panel — each LoRA dropdown fills in the full LoRA list only when you open it (with ~450 LoRAs every click in the panel is ~70 ms faster; nothing changes in how it is used). 总控台更轻：LoRA 下拉框点开时才填完整列表（450 个左右的 LoRA 时，总控台每点一下快 70 毫秒左右；用法不变）。
  • 0.4.2 / 0.4.1: window mode is no longer marked beta; on a ComfyUI older than 0.44, ticking Seamless long video now stops with a clear message instead of an import error. 窗口模式不再标"测试版"；ComfyUI 低于 0.44 时勾"无接缝长视频"会给一句清楚的提示，不再是导入报错。
  • 0.4.0 CCW (Chained Context Windows / 接续窗口): Seamless long video — one switch on the Director: continued shots are rendered together on one timeline (no continuation seams, no drift); hard cuts stay clean; long plans are grouped only at hard cuts. ComfyUI's Context Windows average the window overlaps; CCW runs the windows in time order and each takes over from the one before, so there are no double images. New H3MD Redraw / Preview node: low-res preview, or redraw a few seconds of a finished video (you may change that shot's prompt / pictures first). Interpolation no longer blends two shots at a hard cut. The example workflows cache frames as 8-bit (half the disk, no visible difference). Measured: a 2 min 20 s music video at 1920×1088 in 2 h 18 min on an RTX 5090. 导演台一个开关无接缝长视频：续接的镜头在一条时间线上一起生成，没有接缝、不漂移；硬切照常。官方 Context Windows 把窗口重叠部分平均，CCW 按时间顺序接手前一个窗口的结果，不会重影。新节点局部重绘 / 低清预览：只跑低清，或只重画成片里的几秒（可以先改提示词、参考图）。插帧在硬切处不再混帧。示例工作流默认 8 位缓存（减半，看不出区别）。实测 2 分 20 秒 MV（1920×1088）RTX 5090 用时 2 小时 18 分。
  • 0.3.7 / 0.3.8 (GitHub / Comfy Registry only): files are read and written only in ComfyUI's folders or folders listed in user/h3md/config.json. 只在 ComfyUI 的文件夹或 user/h3md/config.json 里列出的文件夹读写文件。
  • 0.3.6: face refine options (detector … feather) stay editable when face_refine is fed by a link (as in the example workflows); no "unsupported accelerator" warning in ComfyUI-Manager. 修脸选项在 face_refine 连线输入时（示例工作流就是这样）不再被灰掉；管理器不再提示"加速器不受支持"。
  • 0.3.5: motion copy (TARA) no longer fails at the end of a video — the shot length that follows the source video is rounded down in motion copy (prompt with motion_transfer:), and rounded up as before everywhere else; the prompt's Duration: line follows the real segment length automatically. 动作复刻（TARA）跑到视频结尾不再报错：跟随原片的镜头长度在动作复刻时向下取整（提示词有 motion_transfer:），其他功能照旧向上取整；提示词的 Duration: 行自动改成实际段长。
  • 0.3.4: fixed one or two cosmetic UI bugs (no effect on use); skeleton (pose) support with the new H3MD Motion Control node; the simple workflow includes the third-party TARA motion-copy nodes (TARA LoRA + ComfyUI-BFSNodes, bypassed by default — the Control Panel's Motion copy button switches them on); the Control Panel's speed LoRA is now the 768 turbo in the model area (Dasiwa / 10ErosMax rows removed). 修复一两个界面美化上的 bug（不影响使用）；添加骨骼图支持（新节点 H3MD Motion Control）；清爽版工作流里视频复刻加了第三方 TARA LoRA 和 BFS Nodes（默认关，点总控台「动作复刻」自动打开）；加速改由模型区 768 负责（删掉 Dasiwa / 10ErosMax 两行）。
  • 0.3.3: Loop takes T8's semantic bridge directly; face refine also takes the plain ComfyUI-SeedVR2_VideoUpscaler (numz); "which segments" can be typed on the Loop. Loop 能直接接 T8 语义桥；修脸也能接普通版 SeedVR2；Loop 上"跑哪几段"可以直接填。
  • 0.3.2: examples are now just H3MD Director simple.json (start here) and H3MD Director.json (full); the demo pack runs only the first segment until you press Clear. 示例只留清爽版和完整版；演示包默认只跑第一段。
  • 0.3.1: Control Panel without the size drop-down (type width / height); lite and alpha examples removed – use H3MD Director simple.json. 总控台去掉宽高快选下拉；删掉精简版和 alpha 示例，用清爽版。
  • 0.3.0 清爽版: new H3MD Control Panel (size, fps / interpolation, face refine, seed, which segments to run, style buttons Balanced / High motion / Motion copy, two LoRA lists – loads the LoRAs itself); simple workflow with models / sampler / save in subgraphs and no rgthree / Crystools / easy-use / LayerStyle / TE-Speed; simplified Director; original ComfyUI-VOSR2 for face refine. On RunningHub, unpack the subgraphs first. 新"总控台"：常用设置一处改，风格按钮一键切 LoRA；"跑哪几段"一个框；清爽版工作流（子图收起模型 / 采样 / 保存，去掉多个第三方节点）；导演台简化；修脸支持原版 VOSR2。RunningHub 上先把子图释放。
  • 0.2.2 / 0.2.1: safer on shared / server ComfyUI (RunningHub etc.); RefMod list no longer depends on another plugin's private helpers. 共享 / 服务器（RunningHub 等）上更安全；RefMod 列表不再依赖别的插件。
  • 0.2.0: export / import a whole plan (prompts, media, all settings) as one zip; drift_correct on the save nodes (long-take colour / sharpness drift); example workflows use the anti-oily-skin LoRA mix. 计划导出导入；保存时漂移校正；示例工作流换成去油光 LoRA 组合。

Videos / 视频

Run online / 在线运行

RunningHub (referral link, 500 free RH coins / 推荐链接，注册送 500 RH 币): https://www.runninghub.ai?inviteCode=aolwbd1n

License

AGPL-3.0-or-later with attribution: credit "Based on H3MD by YJ" with a link when redistributing or integrating.

Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support