Tensor Binary Op

Tensor Binary Op is where the tensor section of this pack starts doing actual math. You give it two things - two tensors, or a tensor and a plain number - pick an operation from a dropdown, and it applies that operation element-by-element across the whole tensor. It's the "everyone's first arithmetic node" for anyone who needs to manipulate images, masks, or latents numerically instead of relying on whatever pre-built node happens to exist.

It's part of Basic data handling by StableLlama, a dependency-free utility pack that wraps everyday Python and PyTorch into ComfyUI nodes. The tensor corner has create, this binary-op node, a unary-op node, and the reshape/slice/permute/join family.

How it works

Under the hood each input gets coerced to a tensor if it isn't one already (torch.tensor() for scalars, passthrough for real tensors), then the chosen operation runs element-wise. The dropdown offers seven:

• add, subtract, multiply, divide - the basic four
• power - raise the first input to the power of the second
• remainder - modulo, the % operator
• floor_divide - integer-style division that rounds down, the // operator

These are the familiar PyTorch operators, not approximations - torch.add, torch.pow, //, and so on.

The inputs that matter

Three inputs, two of which are the actual operands:

• a (wildcard *) - first tensor or scalar.
• b (wildcard *) - second tensor or scalar. Note that a and b don't have to be the same type: a tensor and a scalar work fine, which is how you'd do "multiply every pixel by 1.2" in one step.
• operation (dropdown, default add) - pick your math.

One wildcard output, ready to wire into anything that wants tensor data.

Where you'd actually use it

This is a workhorse for people doing numerical work on images and masks. Multiply an image tensor by a brightness factor. Add two images to blend them. Subtract a mask from another mask to carve out a region. It's also how you'd experiment with latent-space tweaks that no dedicated node covers - because if an operation is expressible as tensor math, this node can do it.

Installing it

# ComfyUI Manager (recommended): search "Basic data handling", install, restart.

# or manually:
cd ComfyUI/custom_nodes
git clone https://github.com/StableLlama/ComfyUI-basic_data_handling
# restart ComfyUI

No dependencies, no model downloads. If your ComfyUI runs, this runs.

Where people get burned

Three classic traps, all grounded in how the underlying operators behave:

Integer division. If both operands are int tensors, divide does true division and can throw on a zero divisor, while floor_divide truncates silently. If you want fractional results, make sure at least one side is a float - images and latents already are.

Divide by zero. Floats give you inf/nan (which then quietly poison everything downstream); integers throw outright. Either way it's a "check your math" moment, not a node bug.

Shape mismatch. Element-wise ops require matching shapes (or a scalar on one side). If your two images have different heights or widths, PyTorch will refuse with a broadcast error - which is a good error, telling you your data isn't aligned before you've baked in a subtle bug.

One more honest note: with a wildcard input, nothing stops you from wiring in a string, and the resulting error will be cryptic. This node does tensor math; feed it tensors and numbers.