CRE8IT Image Sizer

You know the drill. You load a workflow, the checkpoint is SDXL, the Empty Latent Image is set to 1920×1080, and the result comes back with doubled hands and tiled wallpaper. The image isn't bad - the resolution is. Every model has a native resolution, and generating way past it is one of the most common causes of broken anatomy in the whole ecosystem. CRE8IT Image Sizer (class ImageSizer) exists so you stop hand-typing bad dimensions.

What it does

You give it two things: your model's native resolution class and an aspect ratio. It hands back concrete pixel dimensions that respect both. The three inputs:

• model_type - pick the value matching your checkpoint: 512 (SD 1.5 class), 768, 1024 (SDXL class), or 1280. This is your native-resolution budget.
• aspect_ratio_width and aspect_ratio_height - the shape you want, as plain integers. 16 and 9 for widescreen, 1 and 1 for square, 2 and 3 for portrait.

The outputs are just Width and Height, both INT. Wire them into the width/height of an Empty Latent Image node and you're done. Pick 1024 with a 16:9 ratio and you get a wide canvas kept inside the SDXL native budget, not something that's going to tile.

Why it matters

Native resolutions per family: SD 1.5 is a 512-class model, SDXL is 1024-class, and pushing an SDXL model straight to 1920×1080 is exactly how you get duplicated anatomy and repeated patterns. The fix everyone lands on is the same: generate at or near native resolution, then upscale afterward. This node is the "near native, right shape" half of that. Set the resolution class for your checkpoint once, and every future canvas is one aspect-ratio entry away.

How to install it

It's a pure-math utility - no models, no Python dependencies, nothing in the README beyond the pack's one-line description. ComfyUI Manager → Install Custom Nodes → search ComfyUI-Cre8it-Nodes, install, restart. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/DonBaronFactory/ComfyUI-Cre8it-Nodes

then restart ComfyUI.

Gotchas

The node can't see which checkpoint you loaded - model_type is on you, and getting it wrong is the classic error. Feed 512 to an SDXL model and you get soft, mushy output; feed 1024 to a 512-class model and you get the broken anatomy you were trying to avoid. It also only sets dimensions; it won't upscale or hires-fix for you, so pair it with a real upscale step after the sampler. And honest note: this is a tiny, barely-published personal pack whose GitHub page wasn't resolving when I checked (404) - it typically shows up bundled inside someone's workflow rather than via search.

If you're constantly tweaking compositions it's a nice little time-saver. If you generate one aspect ratio forever, just memorize the numbers and skip it.