One workflow for everything MiniMax H3 can generate.

Full walkthrough on YouTube:
https://youtu.be/yUQ3oa3jULU

H3 makes video and audio at the same time, and it ships as two different UNets
with two different conditioning nodes. This graph has all seven modes in it, so
you don't need a separate workflow per job:

  01  Text to Video          prompt only
  02  Image to Video         one image as the first frame
  03  First / Last Frame     image A to image B
  04  Reference to Video     up to 5 images + 3 videos + 3 audio clips
  05  Motion Tracking        DWPose skeleton
  06  Motion Tracking        DensePose body-surface map
  07  Motion Tracking        SCAIL / NLF 3D mesh

Lip sync is separate from the mode list and works on top of any of the seven.

To use it: pick a UNet in the (1) Model panel, pick a mode in the (2) panel,
hit run. Everything else is optional. The notes inside the graph cover the
model downloads, the resolution table, prompt format and the speed settings.

Full walkthrough on YouTube:
<<ここに YouTube のリンク>>

Two files are attached. The graphs are identical - only the language of the
notes and group labels differs. Take whichever one you read faster.

  *_en_v1.json   English
  *_ja_v1.json   日本語

Custom nodes: rgthree-comfy, ComfyUI-KJNodes, ComfyUI-VideoHelperSuite,
WhatDreamsCost-ComfyUI. Modes 05-07 also need the DWPose / DensePose / NLF
preprocessors. Model download links are in the notes inside the graph.

Please don't change the group colours - the panels find what they control by
colour, so a recoloured group drops out of its panel.