Load Image Batch From Zip

Every ComfyUI user has the same breaking point: you've got 200 reference images, an img2img batch to run, or a dataset you're about to train on, and the only way in is dragging files in one at a time or poking around with the built-in Load Image node. Load Image Batch From Zip is the small node that fixes that exact annoyance. Drop a .zip into ComfyUI's input folder, point the node at it, and out comes a proper batched IMAGE tensor plus a mask - no extraction step, no twenty Load Image nodes, no naming a folder and praying the load order makes sense.

It ships in the ComfyUI-ZipBatchLoader pack, a small single-maintainer project from Charlweed. Don't expect fame: it's a utility, not a headline feature, and that's exactly why you want it in your back pocket for batch work.

How it works

The whole thing is one node with a refreshingly simple job. When the node opens, it scans ComfyUI/input for .zip files and fills the dropdown. When you run it, it opens the archive entirely in memory - no extraction to disk - filters to .png, .jpg, and .jpeg, skips __MACOSX/ junk and hidden dotfiles, and sorts everything alphabetically by filename. Each image becomes two tensors: the RGB version normalized to 0–1 (IMAGE, shape [B, H, W, 3]) and a grayscale version (MASK, shape [B, H, W]), all stacked into one batch. Zero-padded names like 000_ref.png, 010_ref.png, 020_ref.png give you a predictable order you can actually reason about.

The mechanism matters for two reasons. First, it's memory-hungry by design: the whole archive gets pulled into RAM as tensors, so a multi-gigabyte zip of 4K stills can hurt on a small machine. Second, the mask is not an alpha channel - it's just the grayscale conversion of each image. Don't wire it into an inpaint pipeline expecting transparency; it's there so mask-based conditioning has something to chew on.

The inputs and outputs that matter

Only two inputs, and you'll touch both:

• zip_file - the dropdown. It lists every .zip in your ComfyUI input directory. There's no file browser; if your zip isn't there, it can't see it.
• heterogeneous_dimensions - default False. Leave it off and every image must be the same size or the node throws a ValueError naming the offender. Flip it on and mismatched images get silently skipped with a warning instead. For training data that's a lifesaver; for strict batches, keep it off so nothing sneaks through wrong-sized.

Outputs: image (the IMAGE batch), mask (the MASK batch), and count (an INT - handy for confirming nothing got dropped, or as a loop bound). The image output plugs straight into a VAEEncode for img2img-style batches, into Save Image for a bulk export, or anywhere else a normal image wire goes. That's the whole point: it emits standard ComfyUI types, so it plays with everything.

Installing it

Boring and painless, which is good. No models to download, no CUDA surprises - dependencies are just torch, Pillow, and numpy, which you already have.

• ComfyUI Manager: search for ZipBatchLoader (or ComfyUI-ZipBatchLoader) in the Custom Nodes Manager and hit install, then restart.
• Comfy CLI: comfy node install ComfyUI-ZipBatchLoader
• Manual:

  cd ComfyUI/custom_nodes
  git clone https://github.com/Charlweed/ComfyUI-ZipBatchLoader.git

  then restart ComfyUI. It appears in the node menu as Load Image Batch From Zip under the image category.

Where people get burned

Three things trip up new users, all of them visible in the source so you can trust them:

1. The dimension error. Most common by far. Mixed-resolution zips fail loudly with heterogeneous_dimensions off - resize your images before zipping, or accept the skip behavior.
2. The dropdown is stale. The file list is populated when the node loads, so if you drop a new zip in input, refresh the workflow or restart before it shows up.
3. The zip must live in input. It won't find your archive sitting on the Desktop, and it errors with FileNotFoundError if the file vanishes between selection and run.

For a "load a bunch of images" problem, this is the one I'd reach for: one node, predictable order, standard outputs, and zero setup beyond a zip. You'll stop extracting archives by hand before the week is out.