📐AGSoft Aspect Ratio

You want a 16:9 canvas around 1024px on the long side, but every time you type the numbers you get 1024×575 and the model silently ignores your aspect and crops - or worse, stretches anatomy because the resolution sits between its native training ratios. 📐AGSoft Aspect Ratio is the boring fix: tell it the ratio, tell it what's fixed, and it hands you exact, multiple-aligned width and height as plain numbers. No sampling, no latents, no AI. It's a calculator that lives inside your graph so the math can't drift.

It's one node in the big AGSoft pack, which is basically a one-author toolbox of 100+ utility nodes (loops, text, video, masks). This one does one job and does it purely in Python - there's no torch, no model, nothing to download. The source is a ratio parser, a couple of multiplications, and a rounding helper.

How it works

You pick a ratio, an anchor, and a rounding rule, and the node derives the other dimension. The base dropdown decides what "1024" actually means:

• width or height - fix that side, derive the other from the ratio.
• longest / shortest - fix the bigger or smaller side regardless of orientation, which is the right choice for "a 2.39:1 shot at ~2048 wide" without caring about landscape vs portrait.
• megapixels - fix the total area instead, so 1.0 MP gives a ~1:1 1024×1024 but a wider frame at the same pixel budget.

Everything then gets snapped to multiple (default 8, set 32 or 64 to match what a given VAE wants) using round, floor, or ceil. That alignment step is the part that actually saves you: latent models are picky about divisibility, and generating off-multiple just gets you silent cropping or duplicated anatomy when you push past native resolution.

The inputs that matter

• preset - 15 presets (1:1, 16:9, 21:9, 9:16, …) plus custom.
• custom_ratio - only read when preset = custom; accepts strings like 16:9, 1.85:1, 2.39:1.
• base + base_value - what's fixed and its value in pixels (or megapixels_value when base is megapixels). Note the pair: change base to megapixels and the base_value field is ignored, and vice versa.
• multiple and rounding - the alignment rules.

Outputs and where they go

You get four: width (INT), height (INT), ratio (FLOAT), and display (STRING - a ready "1920×1080 (16:9)" line you can pipe into a Show Text or into a filename). These are plain numbers, so wire them anywhere that takes an INT/STRING: a resize node, a crop, a text overlay. AGSoft's own 🧊Empty Latent Plus has this same size engine built in, so if you're using that you don't need this node - reach for the standalone version when you want dimensions for something other than generating, like preparing a consistent canvas across a whole batch.

Installing it

It's part of comfyui-AGSoft, installed once for the whole pack:

cd ComfyUI/custom_nodes
git clone https://github.com/Art-xmaster/comfyui-AGSoft

then restart ComfyUI - or search comfyui-AGSoft in ComfyUI Manager and click install. The pack declares numpy, opencv-python, and translators in its requirements, but this specific node doesn't touch any of them; you're not pulling in anything heavy to use it.

Gotchas

Keep custom_ratio in W:H form - the parser accepts x, /, ,, ; as separators, but a malformed string raises a clear ValueError. And don't set multiple to 1 thinking you'll get exact values: that's the "no rounding" mode, but the node still clamps results up to at least the multiple, so tiny anchors like 10px become 8 or 16. For normal 512–2048 work, default 8 with round is exactly right.