OpenCV getOptimalDFTSize_0

OpenCV getOptimalDFTSize_0 answers one narrow question: "what's the fastest size to run a Fourier transform on that is at least this big?" It wraps cv2.getOptimalDFTSize(vecsize) - give it an integer, get back the smallest integer ≥ that value whose prime factors are only 2, 3, and 5. Those are the sizes OpenCV's FFT handles fastest, and that's the whole job. It's a helper for the pack's dft/idft nodes, and honestly the niche-iest useful node in this family.

Why FFT sizes matter

The discrete Fourier transform is dramatically faster when the input length factors into small primes (2, 3, 5) than when it's an awkward prime like 1021. That's why cv2.dft recommends you pad your data to an "optimal" size first - you lose a few rows/columns of zeros to gain a huge speedup on the transform itself. This node is the calculator for that padding. If you're building any frequency-domain pipeline - band-pass filtering, de-fringing, periodic-noise removal, correlation-based alignment - you'll hit a moment where you must decide what size to pad to. This is the decision, automated.

Inputs and output

• vecsize (INT) - your natural signal length (a row/column count, an image dimension, whatever you're about to transform).
• Output int (INT) - the optimal transform length. OpenCV returns 0 for inputs ≤ 1, and for a vector size that's already optimal it returns the size itself.

Nothing else. No images, no composites, no literal-string traps - this is one of the friendlier nodes in the pack precisely because it's a single integer in and a single integer out.

How it fits a workflow

The real pipeline would look like: decide your transform dimensions with this node, pad your NPARRAY data accordingly (via any padding node or the pack's own tools), run dft, do your frequency-domain surgery, then idft back. In practice, for the majority of ComfyUI users, the honest answer is "you won't touch this" - frequency-domain work on images is a deliberate, niche pursuit, and the pack's dft node works fine on unpadded input, just slower. But if you're doing serious Fourier work, this is the correct first step, and it's the difference between a transform that takes a blink and one that takes a noticeable beat on 4K imagery.

Install

Same as all pack nodes - ComfyUI Manager, search opencv-comfyui (display "OpenCV"), or:

cd ComfyUI/custom_nodes
git clone https://github.com/geroldmeisinger/opencv-comfyui
pip install opencv-contrib-python

Restart, no models. Keep the pack's standing rules in mind for the surrounding pipeline: dft inputs and outputs are NPARRAY (OpenCV numpy), batch size must be 1, and composite types elsewhere still need bracket-literal syntax. For this node itself, there is genuinely nothing that can go wrong - which, in a pack whose README opens with "Expect dragons!", is a small gift.