Extensions/Hongs-node
ComfyUI Extension

Hongs-node

Compatibility fixes and reusable workflow utility nodes for ComfyUI

By hong87LH·Created about a month ago·Updated about a month ago· 1
hong87LH/Hongs-node
Nodes1
On cloudLocal install
CategoryHongs Nodes/MiniMax H3
Stars1
Updatedabout a month ago

Nodes (1)

H3 Turbo Mixed Reference Fix (3-to-2)
Hongs Nodes/MiniMax H3
Readme

Hongs-node

A maintainable ComfyUI custom-node pack for workflow compatibility fixes and reusable routing utilities.

Included nodes

H3 Turbo Mixed Reference Fix (3-to-2)

Use this node between MiniMaxH3ReferenceToVideo and BasicGuider when all of the following are true:

  • MiniMax H3 is running in Turbo/Fast mode.
  • The workflow contains a visual reference (image or video).
  • The workflow also contains an independent audio reference.
  • The Turbo adapter reports a tensor size mismatch such as 3 versus 2.

The node aligns the visual-reference and audio-reference condition times. It does not modify source media and does not patch ComfyUI or the third-party Turbo extension.

Recommended value: shared_reference_time = 1.0.

Installation

Clone the repository into ComfyUI's custom_nodes directory:

cd ComfyUI/custom_nodes
git clone https://github.com/hong87LH/Hongs-node.git Hongs-node

Restart ComfyUI after installation. No additional Python packages are required.

Workflow connection

MiniMaxH3ReferenceToVideo
  -> H3 Turbo Mixed Reference Fix (3-to-2)
  -> BasicGuider

Only enable this fix for Turbo/Fast mixed visual-and-audio reference generation. Quality/original mode can connect the conditioning directly to BasicGuider.

Verified reference combinations

The following MiniMax H3 Turbo/Fast combinations were verified with real generation at 360P, 3:4, 3 seconds, and 6 steps:

| References | Result | |---|---| | 1 image + 1 audio | Passed | | 2 images + 1 audio | Passed | | 1 video + 1 audio | Passed | | 1 image + 1 video + 1 audio | Passed | | 1 image + 2 audio files | Passed |

Routing tests also verify that the fix is not inserted for image-only, audio-only, or Quality/original-mode requests.

Video reference combinations consume substantially more time and memory than image-only combinations. A Fault failed: 2 raised inside the MiniMax audio VAE after several consecutive heavy jobs is a model paging/offload state failure, not the 3-to-2 tensor mismatch. A clean ComfyUI restart recovered the tested tri-modal job.

Maintenance layout

  • nodes/: one module per feature area.
  • __init__.py: stable node registration names.
  • CHANGELOG.md: user-visible changes.

Keep the internal registration key stable after publishing because saved ComfyUI workflows store that key.