Aspect Ratio for 2step Latent Sample

There's a class of fast-sampling workflows - turbo models, two-step latent sampling, that family - where you care a lot about getting the latent dimensions right and very little about doing long division. Aspect Ratio for 2step Latent Sample exists for exactly that: pick a base resolution and an aspect ratio, and it hands you width and height that are divisible by 16, ready to feed into your latent sampler.

It's a small, honest utility node from the AAA Metadata System pack. The description on the tin says it all: "Calculate width and height for 2-step latent sampling based on aspect ratio and base resolution. All outputs are divisible by 16." If you've ever run a latent sampler with a 1080×1000 resolution and gotten a tiling artifact or a "latent shape mismatch" error, you know why the divisibility matters - latent spaces want clean multiples, and 16 is the safe floor.

How it works

You choose a base_resolution (the shorter side, 256–1024 in 64-step increments) and an aspect_ratio from a list of presets - portrait, square, landscape, plus the ultrawide and A-paper ratios. The node computes the other dimension, then rounds both down to multiples of 16. The ratio presets are stored as simple width:height pairs (9:16, 3:2, 16:9, 21:9, and so on), so it's pure arithmetic - no model, no network.

The nice touch: if you connect an image, its actual aspect ratio is used instead of the dropdown, and the dropdown is ignored. That's the workflow for "match this reference image's framing" - drop the reference in, read out the numbers, feed them to your latent node.

Inputs and outputs that matter

• base_resolution - the shorter-side target. 512 is a sane default for SD-family models; go higher for hires work.
• aspect_ratio - the framing preset. Ignored when an image is connected.
• image (optional) - a reference image whose ratio overrides the dropdown.

Outputs: width and height (INTs - wire straight into your latent/sampler sizing inputs), ratio_string (like "16:9", handy for display), and ratio_float (the numeric ratio, if some node wants the raw number).

Installing it

Part of AAA Metadata System by Eric Hiss (GitHub: EricRollei). Via ComfyUI Manager (search "AAA Metadata System") or:

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

Restart ComfyUI. No dependencies beyond the pack's core install.

Common issues

Rounding is the thing to watch. The node rounds to multiples of 16, so a 16:9 at base 512 gives you 912×512, not 911.11×512 - that's correct behavior, but if you're pixel-perfecting a final image, remember the actual ratio is the rounded one, not the ideal one. Some samplers are happier with multiples of 64; 16 is the guarantee here, so for ultra-strict workflows you may still round up yourself. And if you connect an image and the numbers look wrong, check the image's orientation - a rotated source will happily hand back a landscape size for a portrait frame. That's on the source, not the node.