Extensions/ComfyUI MiniMax H3 Latent
ComfyUI Extension

ComfyUI MiniMax H3 Latent

Save and load MiniMax H3 AV NestedTensor latents, avoiding the built-in Save Latent limitation for H3 video/audio latent pairs.

By JerryZRic·Created 3 months ago·Updated 19 days ago· 12
JerryZRic/comfyui-minimax-h3-latent
Nodes2
On cloudLocal install
Categorymodel/latent/minimax
Stars12
Updated19 days ago
Readme

MiniMax H3 Latent

English | 简体中文

ComfyUI custom nodes for saving and loading MiniMax H3 AV latents.

MiniMax H3 latents are not plain tensors. They are NestedTensor values containing video and audio latent members, so ComfyUI's built-in Save Latent node can fail with:

'NestedTensor' object has no attribute 'contiguous'

This node pack adds H3-specific latent save/load nodes.

Nodes

  • Save MiniMax H3 AV Latent

    • Saves H3 LATENT inputs, including NestedTensor video/audio members.
    • Files are written to the ComfyUI output folder with the .h3latent extension.
  • Load MiniMax H3 AV Latent

    • Loads .h3latent files from the ComfyUI input and output folders, including subfolders.
    • Output files have an [output] suffix in the file selector; existing input selections remain compatible.
    • Rebuilds the H3 NestedTensor before returning a LATENT.

Usage

Use Save MiniMax H3 AV Latent instead of the built-in Save Latent node for MiniMax H3 workflows.

To load a saved latent, use Load MiniMax H3 AV Latent and select it directly from the output folder, for example latents/MiniMaxH3_00001_.h3latent [output]. There is no need to move it to the input folder. Files already in the input folder are still available without an [output] suffix.

Refresh the node definitions using ComfyUI's refresh command (or reload the browser page) after saving a new file to update the file selector. Overwriting a selected file changes its content hash, so it will be loaded again on the next execution. The save node's existing numbered filenames are unchanged; this update does not add automatic overwriting.

Format

The saved file stores each latent member separately:

latent_0  video latent
latent_1  audio latent

On load, files with multiple members are rebuilt as a ComfyUI NestedTensor.