Load ZIP file

The usual way to get a folder of images into ComfyUI is a small ceremony: unzip on your desktop, copy the folder into ComfyUI/input, refresh, and hunt for the right Load node. This node skips all of it. Drop a ZIP file onto a button inside the node and it reads the archive straight into the graph - images, masks, captions, video and audio at once, each as its own typed output. It's the centerpiece of this pack and the reason the pack exists.

How it works

The upload button is a real browser upload: it POSTs the file to ComfyUI's /upload/image endpoint, and the archive lands under ComfyUI/input/uploaded_zips/. On RunPod or Colab the same thing happens - your browser pushes the ZIP to the remote server, so there's no ssh/drive dance for getting data in.

On execution the node walks every member of the archive and sorts it by extension:

• Images - png, jpg/jpeg, webp, bmp, tif/tiff
• Video - mp4, webm, mkv, avi, mov, m4v, gif
• Audio - wav, mp3, flac, ogg, m4a, aac
• Text - txt, json, srt, vtt, csv, md, log

That file is not decompressed to input - members are read in memory (video gets extracted to a temp dir and becomes a proper ComfyUI VIDEO object).

The pairing conventions that make it a dataset node

This is where it stops being "load a folder" and becomes genuinely useful. The pack uses filename stems to pair things up, the same way a training dataset is laid out on disk:

• cat.png + cat_mask.png in the same ZIP → the mask gets loaded and wired to the MASK output.
• cat.png + cat.txt → the txt becomes that image's PROMPT WORD (and also shows up raw in the TEXT output, since all .txt files are loaded).
• A PNG that carries an alpha channel with no _mask sibling gets its alpha turned into a mask (1 − alpha, so transparent areas read as mask=1).

Then the sequences are aligned by stem and zipped together, so batch element i of IMAGE, VIDEO, AUDIO and TEXT all come from the same logical slot. That alignment is the whole point: feed the outputs to a sampler and each element is one job in the batch.

The outputs that matter

You'll use IMAGE, MASK, PROMPT WORD and TEXT most. Two outputs are plumbing:

• Naming - a dict of the original filenames inside the ZIP. Wire it into Save to ZIP's naming input and your original file names come back out the other end.
• batch_cursor - an accumulator. Feed it into a second Load ZIP node's batch_cursor input and you can chain several archives into one combined batch. The load toggle (default on) controls whether a node reads its ZIP; flip it off for a pure pass-through.

Installing

ComfyUI Manager → Install Custom Node → search for jayhuang-zip-unpack-images, or:

cd ComfyUI/custom_nodes
git clone https://github.com/aixqueen/jayhuang-zip-unpack-images.git

Then restart ComfyUI. No models to download. The one dependency is soundfile, and it only matters for non-WAV audio in your ZIPs - WAV decodes with the Python stdlib. If you load an mp3 or flac and hit a "Missing 'soundfile' package" error, that's it:

pip install soundfile

Gotchas

The node re-runs whenever the ZIP's size or mtime changes, so re-uploading a same-named archive does trigger a fresh read - that's a feature for iteration, but it means a zip sitting on a synced drive can fire the node when you least expect it. A corrupted archive fails with a clear "ZIP file is corrupted" error, and if the archive contains none of the supported extensions it raises "No supported content found."

One honest caveat: this is a tiny, single-commit pack from an author with effectively no community footprint. The mechanism is straightforward and reads cleanly, but with ComfyUI custom nodes you're executing arbitrary Python - give it the same glance-over you'd give any unknown node, then enjoy the part where your workflow finally stops treating a folder as a problem.