MD: Empty Latent Ratio Select

If you've ever hand-typed "1216x832" into an Empty Latent Image and crossed your fingers, this node is for you. MD_EmptyLatentRatioSelector does what its name says: it generates a latent at a chosen aspect ratio and megapixel target, with the width and height computed to land on multiples of 8 so nothing gets snapped weirdly downstream.

It's a small utility, and it earns its place because getting resolution right is one of the cheapest wins in ComfyUI. Models are trained at a specific pixel count - SDXL around 1.0 MP, SD1.5 around 0.26 MP, FLUX happier a bit higher - and wandering far from that is the classic cause of doubled-up subjects, composition weirdness, and tiling artifacts. This node keeps the total area on target while letting the shape float.

How it works

You pick a ratio, an orientation, and a total megapixel budget. The node computes width = sqrt(area × ratio) and height = area / width, then rounds both down to a multiple of 8 (the CONST_DIVISIBLE_BY = 8 in the source) before building the latent via ComfyUI's empty latent constructor. Portrait just swaps the axes. The math keeps total pixels near your target, so a 21:9 output isn't secretly a 16:9 crop with extra pixels.

The inputs you actually set

• ratio_preset - 1:1, 16:9, 21:9, 32:9 and the in-between ratios. The tooltip's warning about extreme ultrawides is real: if your model wasn't trained on them, you'll see duplicate subjects.
• orientation - landscape or portrait. Match it to the prompt's intent (portrait for characters, landscape for environments).
• base_mp_preset - the resolution class: SD1.5, SDXL, 1080p, 4K, or Custom. Match this to the model you're loading. Pushing SDXL to 8.3 MP without an upscale stage is how you get artifacts, not detail.
• manual_mp_size - only used when the preset is Custom; a slider from 0.1 to 16 MP.
• dimension_scale - a linear multiplier for the final dimensions. 0.5 gives you a half-res test render without touching the MP logic. Handy for rapid iteration.
• batch_size - how many latents in one batch; VRAM scales with it.

Outputs: the latent (into your KSampler), plus width and height as INTs so you can feed resolution into a filename builder or a text node if you want.

Installing

It ships with the MD Nodes pack. Either way you do it - ComfyUI Manager (search MD Nodes) or a manual clone - the steps are the same:

cd path/to/ComfyUI/custom_nodes
git clone https://github.com/MDMAchine/ComfyUI_MD_Nodes.git
cd ComfyUI_MD_Nodes
pip install -r requirements.txt

Then restart ComfyUI. This particular node is pure Python with no extra dependencies, so if you only care about this one, you can even skip the requirements install and it'll still load - but the pack's audio nodes will complain until you install their deps.

Common gotchas

The biggest one is expecting more than it delivers: this is a resolution calculator, not an upscaler. A 4K preset at native SDXL resolution doesn't magically give you 4K detail - it gives you an SDXL-sized image stretched to a big latent, which often looks worse than generating at native size and upscaling. Keep dimension_scale at 1.0 for finals and use the low presets for testing.