Image to SDXL compatible WH

SDXL has a quirk that trips up everyone coming from SD 1.5: it was trained at 1024x1024 with a fixed set of aspect ratios, and the model is genuinely picky about it. Generate at some arbitrary resolution and you get artifacts, composition drift, or just a model that's clearly off its game. The accepted practice is to pick the closest native bucket - 1024², 832x1216, 896x1152, and friends - instead of any resolution you feel like. Image to SDXL compatible WH does the picking for you: feed it any image, it measures your aspect ratio and returns the closest SDXL-native width and height. You feed those numbers into an EmptyLatentImage or a resize, and your img2img / ControlNet work suddenly behaves.

How it works

One input: image. The node reads the image's aspect ratio (width ÷ height), then finds the closest match from the fixed table of SDXL training buckets - all 13 of them, covering 1:1, 2:3, 3:4, 5:8, 9:16, 9:19, 9:21 and their landscape mirrors, each at its native pixel count. It returns two INTs: width and height.

Note what it does not do: it doesn't resize, crop, or pad anything. It's a calculator with one job, and it does it because an aspect-ratio-aware calc is exactly what keeps people from hand-rolling 1216x832 and missing the bucket by a hair. The output numbers are the whole product.

How to actually use it

In an img2img or ControlNet workflow: take the image you're conditioning on, run it through this node, and wire the two INTs into an EmptySDXLLatent (or a resize-to-size node) that your sampler consumes. That way the generation canvas is always a true SDXL bucket that matches your source's composition - the standard recipe for "condition on this image without mashing the aspect ratio." If you instead want the condition image itself resized to the bucket, chain the INTs into a resize node before the encoder. The node itself is agnostic about which you do; it just makes the numbers easy.

The nearest-bucket logic is adapted from a community snippet (credited in the source), and the pack keeps a copy so you don't need a separate utility pack just for this. It's dead simple, deterministic, and exactly the kind of thing that's faster to look up than to get wrong.

Install

Ships in ComfyUI-GlifNodes from the glif.app team. ComfyUI Manager → "ComfyUI-GlifNodes", or:

cd ComfyUI/custom_nodes
git clone https://github.com/glifxyz/ComfyUI-GlifNodes

Restart ComfyUI. No models, no dependencies, instant - it's an aspect-ratio table and a min() call.

Where it shines: SDXL img2img, Redux/IP-Adapter conditioning that needs to match source composition, and batch workflows where every image should be snapped to a native bucket before the sampler sees it. If you only use SDXL occasionally, this is the reminder you'll be glad to have.