SDXL Resolution Presets (ws)

If you've ever typed 1920, 1080 into an Empty Latent Image node for an SDXL checkpoint and wondered why you got a second head growing out of someone's shoulder, this node exists for exactly that reason. SDXL wasn't trained at arbitrary resolutions - it was trained at 1024x1024 and a handful of other fixed aspect ratios, and stepping outside that set is one of the most common ways people accidentally sabotage their own generations. SDXL Resolution Presets just gives you a dropdown of the ratios that actually work, so you can stop doing width/height math in your head and stop guessing.

What it actually is: a tiny, single-purpose picker. It's not a resizer, not an upscaler, not touching your image at all - it's a lookup table wearing a ComfyUI node costume. You pick a label, it spits out two numbers.

How it works

Under the hood this is basically the resolution-selector idea from twri's sdxl_prompt_styler pack (the README credits it directly), stripped down to just the width/height part. The dropdown has eight options - Square (1024x1024), two Cinematic orientations (1536x640 / 640x1536), two Widescreen orientations (1344x768 / 768x1344), two Photo orientations (1216x832 / 832x1216), and Portrait (1152x896) - and every single one of those numbers matches SDXL's official trained aspect-ratio bucket, including the rotations. That's not a coincidence; it's the whole point. Stability trained the base model on a fixed set of buckets close to 1024², and generating at one of them is what keeps anatomy coherent. Go off-ratio and you get stretching, duplicated limbs, or the classic "two heads" artifact, because the model is trying to fill composition space it never saw examples of. This node is just a guardrail against that mistake, wearing a friendly UI.

Inputs and outputs

There's exactly one input that matters: resolution, an enum dropdown defaulting to Square (1024x1024), with the eight presets above as the only choices. No custom values, no sliders - you're choosing from a fixed menu, which is the feature, not a limitation.

It gives you two outputs: width and height, both plain integers. Wire those straight into the width/height sockets on your Empty Latent Image node (or anywhere else that wants raw pixel dimensions - an upscale node, a resize calculation, whatever). That's the entire workflow: dropdown in, two numbers out, numbers into your latent.

Installing it

This lives inside the hylarucoder/comfyui-copilot pack, which per its own README is explicitly "self-used nodes" - a solo developer's personal toolbox, not a polished, widely-registered community pack. Don't expect it to show up reliably as an auto-install suggestion when you load someone else's workflow with a missing-node warning. The safe route:

cd ComfyUI/custom_nodes
git clone https://github.com/hylarucoder/comfyui-copilot

Then restart ComfyUI. Try ComfyUI Manager's search first (it's registered under the pack title "comfyui-copilot"), and fall back to the manual clone if Manager comes up empty. There's nothing to download beyond the code - no model weights, no checkpoints, and the README lists no special Python dependencies, so this is about as low-risk an install as custom nodes get. It won't drag you into the dependency-conflict mess that plagues heavier packs.

Common issues

The main gotcha isn't really a bug, it's a naming trap: search "comfyui-copilot" and you'll likely also surface a much bigger, unrelated project that shares almost the same name - an AI chat assistant that helps you build ComfyUI workflows. That's not this. This is a two-node personal utility pack from a different author entirely; if Manager or a search engine points you at something with a chat interface, you grabbed the wrong repo.

Beyond that, the only real limitation is that the menu is the menu - if you need a resolution outside these eight (say, a 21:9 ultrawide, or a resolution tuned for a specific SDXL fine-tune trained on a broader range), this node simply can't give it to you. Skip it and wire your own numbers into Empty Latent Image directly. And a quick sanity check worth keeping in your head: these presets are the SDXL trained set specifically. If you're running an SD 1.5 checkpoint or something with a genuinely different native resolution, this node's outputs are still valid numbers, just not the ones tuned for that model - pick with your actual base model in mind, not just the node that happens to be sitting in your library.