Image math

Image math runs a math expression over every pixel of up to four images at once. The default expression, a*(1-w)+b*w, is a weighted blend - which means this single node replaces the "Image Blend" node you'd otherwise add, with the bonus that you can write a*b, abs(a-b), 1-a, or whatever your effect actually needs instead of hunting for a node that does it. It's the same idea as WAS Node Suite's image math, delivered as a programmable expression rather than a fixed set of blend modes.

How it works

In ComfyUI an image is a (batch, height, width, channel) float tensor, values 0–1. This node evaluates your expression elementwise on that tensor. The expression language gives you more than the four inputs: inside it you have a, b, c, d (the images), w, x, y, z (float sliders), and positional variables - X/Y (pixel coordinates), W/H (dimensions), C (channel), B (batch), T/N (batch count / channel count). So you can build gradients from coordinates, or grade a single color channel with C.

The function library extends into image territory too: blur, edge, dilate, erode, morph_open, morph_close are all callable inside the expression, so feathering or a quick edge pass doesn't need a separate filter node.

The inputs that matter

• a - required image. b, c, d optional.
• Image - the expression. Default a*(1-w)+b*w.
• length_mismatch - how to handle different batch sizes: tile (repeat), error (default, fail loudly), pad (zero-fill).
• w, x, y, z - the float knobs your expression uses.

Output is a single IMAGE.

Installing it

Part of More Math (mcDandy/more_math). Fastest: ComfyUI Manager, search "More Math". Or:

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

Then restart. Dependencies are antlr4-python3-runtime and torch - no model files. The pack uses ComfyUI's newer node API, so keep ComfyUI updated or the nodes simply won't register.

Where people get burned

The big one: pixel values must stay in 0–1. Write a+b on two full-brightness images and you get blown-out whites because 1+1=2, and a viewer will clip or just show you garbage. Wrap it: clamp(a+b, 0, 1). clamp is your friend here, and a * for intersection or a 1-a for inversion are the operations you'll use most.

Second, length_mismatch defaults to error. Feeding images with different batch counts (say a 1-frame image against a 24-frame batch) throws a ValueError instead of silently doing something weird - which is good, but you'll hit it the first time you wire a single reference image into a video expression. Switch to tile when you mean "repeat this one across the batch."

Also note: like most of the pack's per-type math nodes, the classic form here is deprecated in favor of an autogrow variant with the same "Image math" display name. Both ship and both work; the newer one just takes more image inputs. If you're writing a new workflow and see two "Image math" entries in the menu, the autogrow one is the future-proof pick.