Extensions/Mask Tensor I/O
ComfyUI Extension

Mask Tensor I/O

Save and load ComfyUI MASK tensors as .pt files

By swanjohn99·Created 22 days ago·Updated 17 days ago· 1
swanjohn99/comfyUIcostumNodes
Nodes7
On cloudLocal install
Categorymask_io, bbox_io
Stars1
Updated17 days ago

Nodes (7)

Build Media IO Basename

Stop naming your bbox, mask and latent files after 00007

mask_io
Load Bounding Boxes

Feed a saved detection pass back in and skip the detector

bbox_io
Load MiniMax H3 AV Latent

Pull a saved H3 latent back in and redo only the decode

h3_latent_io
Load Mask Tensor

Get yesterday's mask back, in full float precision

mask_io
Save Bounding Boxes

Cache the detection pass before you re-render anything

bbox_io
Save MiniMax H3 AV Latent

Bank the sampler output so a 33B video render isn't a one-shot

h3_latent_io
Save Mask Tensor

Stop re-segmenting the same video every time you tweak a denoise

mask_io
Readme

ComfyUI Mask Tensor I/O

Save and load ComfyUI MASK tensors as .cmask files (exact float values).

Nodes (category mask_io):

  • Save Mask Tensor - write MASK to output/masks/{prefix}_{#####}.cmask, pass-through
  • Load Mask Tensor - read a .cmask (or legacy .pt) back as MASK [B,H,W]
  • download (both nodes) - opens a picker of output/{subfolder}/*.cmask; download and/or delete selected files (survives ComfyUI restart)

Legacy .pt mask files still load, but ComfyUI may false-flag them as missing models. Re-save with Save Mask Tensor and switch the workflow to the new .cmask file.

Install

cd ComfyUI/custom_nodes
git clone <this-repo-url>

Restart ComfyUI. No pip install; uses Comfy's torch and folder_paths.

Comfy Registry / Manager

  1. Create a publisher + API key at https://registry.comfy.org (set PublisherId in pyproject.toml if different from swanjohn99).
  2. Publish: comfy node publish
  3. Install from ComfyUI Manager, or search the registry.

Usage

  1. Wire SAM3_TrackToMask masks into Save Mask Tensor, then into your mask consumers.
  2. Later: Load Mask Tensor instead of re-running SAM3.