🔧 SDXL Resolutions

SDXL has opinions about image size, and if you ignore them you get duplicated heads, stretched bodies, and mushy composition. The model was trained with aspect-ratio bucketing - a fixed set of width/height pairs that all land around one megapixel - and it behaves best when you generate at one of those blessed sizes. SDXLResolutionPicker is just a dropdown of those sizes that spits out the width and height as numbers, so you stop hand-typing 832×1216 and getting it slightly wrong.

It's a small quality-of-life node from ComfyUI Essentials by cubiq (Matteo Spinelli, also the author of the widely used ComfyUI_IPAdapter_plus). Nothing clever under the hood - the value is that someone curated the list so you don't have to memorize it.

How it works

The single control is a dropdown of 26 resolutions, each labeled with its aspect ratio, from tall portrait (704x1408 (0.5)) through square (1024x1024 (1.0)) to wide landscape. Pick one and the node outputs the two integers. That's the entire mechanism - it's a lookup table with a nice UI, and the reason it exists is that "generate at a native SDXL bucket" is advice everyone gives and nobody wants to keep a cheat sheet for.

The inputs and outputs that matter

• resolution (dropdown, default 1024x1024 (1.0)) - the only thing you set. The number in parentheses is the width÷height ratio, which is handy when you're matching a target frame.
• width (INT) and height (INT) - wire both into your Empty Latent Image node. That's the standard hookup: picker → empty latent → sampler.

How to install it

ComfyUI Manager: search ComfyUI Essentials, install, restart. Or by hand:

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

and restart. No models, no heavy dependencies - it's a list of numbers.

Common issues & troubleshooting

It didn't make my image sharper. It won't. This node picks a good resolution; it doesn't upscale or add detail. If you want a bigger, crisper output, generate at a native size here and then send it through a proper upscale pass afterward.

These are SDXL sizes - I'm on Flux/SD1.5/something else. The bucket list is tuned for SDXL's ~1MP training. SD 1.5 wants sizes around 512–768, and Flux is far more forgiving about resolution, so this particular list is the wrong cheat sheet for those models. Use it for SDXL and its descendants (Pony, Illustrious, and the like), where the buckets genuinely matter.

Off-bucket sizes were "working" before. They can look fine at ratios close to a bucket, which is why the failure is sneaky - push the aspect ratio far from a trained size and that's when SDXL starts cloning subjects and losing coherence. Staying on the list is cheap insurance.

The node went missing after an update. Essentials has been maintenance-only since April 2025, so newer ComfyUI releases occasionally break pack nodes. Update through Manager; if it's genuinely incompatible, this one is trivial to live without - any two integer nodes feeding your latent will do the same job.