Skip to content

Latest commit

 

History

55 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

ComfyUI MiniMax H3 Guide

A dependency-free ComfyUI node pack for building MiniMax H3 prompts from explicit references, subjects, and shots. It compiles those choices into valid H3 prose and routes the media to ComfyUI's native H3 nodes.

  • Typed Plan v2 workflow instead of one large prompt form
  • Explicit identity, keyframe, motion, edit-source, voice, and soundtrack roles
  • Deterministic labels, timing, validation, and native media routing
  • Optional Qwen enhancement with compiler-owned structure kept locked
  • Reusable image/audio Reference Sheets and a locked-picture Foley mode
  • H3-native AV split/repack with frame-locked audio length

Install

cd ComfyUI/custom_nodes
git clone https://github.com/ethanfel/ComfyUI-MiniMax-H3-Guide

Restart ComfyUI, then look under MiniMax H3 → Plan v2. The pack has no additional Python dependencies.

This release requires a ComfyUI build with native MiniMax H3 support. Update ComfyUI if MiniMax H3 Image to Video, MiniMax H3 Reference to Video, or the MiniMax tokenizer is missing.

Start here

Project Setup
  → optional Image / Video / Audio References
  → one Shot per real cut
  → Prompt Merge
  → Apply Reference Plan
  → sampler
  1. Project Setup selects the output size and native frame count.
  2. Reference nodes declare what each file means. The compiler assigns H3 labels and routes automatically.
  3. Shot nodes define the timeline. Attach keyframes, motion, and dialogue to the Shot that uses them.
  4. Prompt Merge validates and compiles the plan.
  5. Apply Reference Plan hands the matching prompt, media, and AV latent to ComfyUI's native H3 implementation.

For the complete node-by-node workflow, see the Plan v2 guide.

Pick roles by intent

Intent Use
Exact opening or ending composition Image Reference as an endpoint
Preserve identity, wardrobe, object, scene, or style Image Reference as reusable content
Transfer pose, action, or body mechanics Video Reference + Attach Motion
Modify an existing video Video Reference as an edit source
Continue an existing video Video Reference as a continuation source
Keep or copy source audio Paired Audio Reference with continuity/copy role
Borrow voice, beat, music style, or sound texture Audio Reference with the matching guidance role

The same file type can take different routes. Its intended job—not its extension—decides the H3 mode.

Example workflows

Examples are available in example_workflows/ and ComfyUI's workflow template browser.

Workflow Status Purpose
Animate Keyframe with Motion Reference Ready Retarget motion onto a supplied opening keyframe
Prompt Builder App WIP Prompt-only Plan v2 interface
Identity and Voice WIP Identity, voice, dialogue, and Ref2VA generation
First and Last Frames WIP Exact endpoint composition
Five Shot Keyframe Composition WIP Per-shot keyframes and motion transfer
Character Replacement WIP Replace one performer in an edit-source video
Video Extension with Audio Continuity WIP Character transfer plus continuation
Video to Audio Foley WIP Locked picture track with generated audio
Foley with Sound Reference WIP Foley guided by a sound texture
Foley with Multiple Sound and Voice References WIP Multi-reference Foley and dialogue

The ready motion-reference workflow is the reference implementation. WIP templates expose their intended graph but still need task-specific media and default refinement.

Limits worth knowing

Constraint H3 limit
Output 4–15 seconds on the native 17k+5 frame grid at 24 FPS
Images 9
Videos 3
Audio clips 3
Mixed reference files 12 total
Each reference video or audio clip 2–15 seconds
Cumulative video references 15 seconds
Cumulative audio references 15 seconds

The 362-frame endpoint plays for 15.083 seconds. One complete-copy or continuity soundtrack paired with a matching 362-frame video may share that padded boundary; it does not increase the general audio limit.

Guides

Legacy nodes remain registered so existing workflows load. New projects should use Plan v2.

Development

pytest -q

Credits

The interaction direction was inspired by ComfyUI-MiniMaxH3-Easy, used under its MIT license. This project retains an independent Plan v2 data model, validator, router, and compiler.

The H3 AV timing nodes were cross-checked against ComfyUI's native H3 grid formulas and the independent findings in ComfyUI-MMH3Tools.

Prompt structure follows MiniMax's official base guide, reference guide, and model card.

License

GNU General Public License version 3 only (GPL-3.0-only).

About

Guided MiniMax H3 prompt preparation node for ComfyUI

Resources

Stars

209 stars

Watchers

2 watching

Forks

Releases

Packages

Contributors

Languages