Nearest Divisible Int

Diffusion models are picky about dimensions. SD 1.5 wants multiples of 8, SDXL and a lot of modern models want multiples of 16 or 64, and video models like Wan get genuinely unhappy with off-grid resolutions. Nearest Divisible Int is the tiny node that fixes it: give it a number and a divisor, and it returns the nearest integer that divides evenly by that divisor. Chain two of them (one for width, one for height) and any arbitrary resolution gets snapped to a safe grid before it hits a sampler.

It's from artyclaw/artyclaw-comfy, and it's the kind of utility that earns its keep silently - you stop thinking about it the moment your "odd" resize results stop producing artifacts or VAE warnings.

How it works

Pure arithmetic, and the source is explicit about the tie-breaking rule. Given value and divisor:

• lower = value rounded down to the nearest multiple.
• upper = lower + divisor.
• If the value is closer to upper, it rounds up; otherwise down. Ties go up - at exactly halfway, you get the higher multiple.

So 514 with divisor 16 → 512 (down), 520 with divisor 16 → 528 (up, since 520 − 512 = 8 ≥ 8, the half-way point). A divisor of 0 or negative is ignored and the value passes through unchanged, so a misconfigured graph fails safe instead of crashing.

Inputs and output

• value - the integer to adjust (default 512, max 999999).
• divisor - the grid you want (default 16, up to 1024). 8 for SD 1.5-era pipelines, 16 for most modern ones, 64 if you're being extra safe on a video model.

One output, result (INT). Typical wiring: put it between a "calculate resolution" node (or a free-text size) and a latent/Empty Latent Image's width input.

Install

ComfyUI Manager → search ArtyClaw Comfy Nodes, or:

cd ComfyUI/custom_nodes
git clone https://github.com/artyclaw/artyclaw-comfy

Restart. Standard library (math) only.

Where people get burned

The tie-goes-up rule surprises people who expect round-half-down - 512 with divisor 16 is already a multiple so it's a no-op, but 520 → 528 (not 512), which can be one step higher than you eyeballed. Second, it takes integers only: feed it a float and you'd better have a conversion node upstream, because this snaps whole numbers. Third, remember it rounds to the nearest multiple, not down - if you have a hard VRAM budget, you may want to round down instead, and this node won't do that for you. Small, predictable, and worth having in the toolbox precisely because it removes a whole class of "why is my output stretched/tiled" questions.