🌀 BV Image Difference Heatmap

"How much did that actually change?" is a question you ask constantly once you start iterating on generations, and eyeballing two images side by side is a terrible way to answer it. BV Image Difference Heatmap gives you the honest version: it compares two images pixel by pixel and paints the differences as a heatmap, plus hands you a couple of hard numbers so you can stop squinting.

The realistic use cases: before/after an upscale or img2img pass, checking how much a low-denoise refiner actually moved, or comparing two generations against a reference. It's the kind of node that's not going to make your images better by itself - it makes your judgments better, which eventually does.

The inputs

• image1 / image2 (both required) - the pair to compare. Feed them same-size images; a 512×512 against a 2048×2048 comparison is meaningless even if it doesn't error.
• overlay_strength (default 0.5, range 0–5) - how strongly the heatmap blends over the originals in the overlay outputs. Crank it toward 5 if the difference is subtle and you want it loud.
• colormap - the palette, straight from matplotlib's set: jet, viridis, plasma, inferno, magma, cividis, turbo, hot, cool, plus the reversed variants of most of them (jet_r, viridis_r, ...). For "where is the difference," jet or turbo are the classic choices; hot reads nicely for magnitude.

The outputs

• heatmap (IMAGE) - the raw difference map in your chosen colormap.
• overlay_image1 / overlay_image2 (IMAGE) - the heatmap blended over each original at overlay_strength. These are what you'll actually want to look at.
• difference_strength (FLOAT) - a scalar summarizing how much the images differ overall. Feed it into a conditional node to gate the rest of the workflow on "did this change enough?"
• mse_value (INT) - the mean squared error, rounded to an integer. The boring-but-useful number for logging or comparing runs.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/BlackVortexAI/ComfyUI-BVortexNodes.git

or ComfyUI Manager → "BV Nodes", then restart. The colormaps come from matplotlib, which ComfyUI usually has on hand - if the node throws an import error, pip install matplotlib fixes it.

Where people get confused

The mse_value output being an INT rather than a float is a minor quirk - don't expect fractional precision from it. And remember the overlay outputs are where the visual answer lives; the raw heatmap alone is harder to read because the dark pixels of the original aren't there as a reference frame. Same-size images in, comprehensible answer out - that's the whole contract.