Calculator

Here's a genuinely useful move with hyperspectral data: pick a region of your image, and ask what the average spectrum of that region looks like. That single question powers a huge amount of remote-sensing analysis - "what does the typical vegetation pixel reflect?", "how does this mineral class's spectrum differ from that one?" The Calculator node is that question, packaged for a ComfyUI graph.

Don't let the generic name fool you. Right now it does exactly one thing, and it does it cleanly.

How it works

You feed it a spectral image (the pack's SPE type, from either loader) plus one or more masks, and it computes, for each mask, the mean of every pixel where the mask is nonzero - averaged across all pixels but not across bands. So each mask becomes one row in the output, and the row has one value per band: that's the region's mean spectrum.

Mechanically it's about as simple as it gets. For every mask in the batch:

raw[mask != 0].mean(0)

The output lands as a 2D NDARRAY with shape (number_of_masks, number_of_bands). Nothing fancy, no learning, just honest statistics.

The inputs that matter

• image (SPE) - the hyperspectral object, straight from Spectral Loader or ENVI Loader.
• masks (MASK) - any ComfyUI mask works, but the intended pairing is obvious from the pack's example workflow: the KMeans node's classes mask output. Run KMeans over your cube to segment pixels into classes, feed a few of those class masks in, and you get the mean spectrum per class - which is the standard unsupervised way to start understanding what's in a scene.
• method - an enum dropdown, and here's the honest part: it only contains mean, and that's its default. The dropdown is future-proofing, a placeholder for methods that don't exist yet. If you somehow picked something else, the code would raise NotImplementedError anyway. So ignore it; it's a promise, not a feature.

Output

results (NDARRAY) - the stack of mean spectra. Its natural destination is the pack's Plot node, whose series input expects exactly this shape: one spectrum per row, plotted against band number. That's precisely what the repo's example workflow does - KMeans segments, Calculator averages, Plot draws.

Installing it

ComfyUI Manager: search ComfyUI_Spectral, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/chenlongming/ComfyUI_Spectral
cd ComfyUI_Spectral
pip install -r requirements.txt

then restart. (The README's reqirements.txt is a typo - the real file is requirements.txt.) The only dependency you probably don't already have is spectral; there are no model files to fetch.

Where people get tripped up

Two things. First, the MASK input isn't a single mask the way you might reflexively expect - it's a batch, and each mask in the batch becomes its own output row. That's a feature: it's how you compute a whole set of class spectra in one pass. Second, keep your expectations calibrated on the method dropdown - this is an early-stage pack (the README says so plainly), so "Calculator" is a name with ambitions, not a spec sheet. One mean-spectrum trick today, more later, maybe.

The real trap isn't the node, it's the data: raw[mask != 0].mean(0) averages everything the mask selects, so a sloppy mask (feathers, partial pixels, mixed classes) gives you a mushy spectrum that's a blend of whatever it overlaps. If your mask straddles two materials, the "mean spectrum" is neither. Clean masks in, clean spectra out.