Select latent size 1Mp

This one's a dropdown of named resolutions sitting right around 1 megapixel - "Square (1024x1024) 1x1," "HD_Video landscape (1280x720) 16x9," "Photo landscape (1216x832) 3x2," and their portrait mirrors, seventeen entries total. That's not an arbitrary size band: SDXL, and by extension every SDXL finetune (Pony, Illustrious, NoobAI), was trained on a set of 1024x1024-area buckets, and several entries here - the square, the 1536x640 cinema size, the 1216x832 photo size - line up with SDXL's own documented trained ratios almost exactly. Pick by name here and you're landing inside the range these checkpoints actually know, rather than typing numbers and hoping.

That matters more than it sounds. Off-ratio, off-scale generation on an SDXL-family model is the classic source of duplicated limbs, stretched anatomy, and repeated backgrounds - the model is extrapolating outside what it was trained to produce. A curated list like this exists specifically to keep you out of that territory.

How it works

Pick a named resolution from the dropdown - landscape and portrait entries are separate, so you don't need a swap toggle, just scroll to the mirrored one. batch_size stacks copies in the output, same as any standard empty-latent node.

Inputs and outputs that matter

• resolution (default Square (1024x1024) 1x1) - the named size to build the latent at.
• batch_size (1–10, default 1) - how many copies to generate in the batch.

One output: Latent, an empty latent at your chosen size and batch count, ready to wire straight into a sampler.

Installing it

ComfyUI Manager: search mzmaxam in Custom Nodes Manager and install. Or by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/MzMaXaM/ComfyUi-MzMaXaM

Restart ComfyUI. No models, no dependencies - clone and restart is the whole install. The pack also ships example workflows under Workflow → Browse Templates → comfyui-mzmaxam.

Where people get tripped up

This node is dropdown-only - there's no free-text width/height field, so if your exact target size isn't listed, you'll want either Select Latent Size Plus (the pack's more flexible sibling, with 148 resolution options and separate width/height outputs) or ComfyUI's stock Empty Latent Image node. It also doesn't output the raw width and height as numbers, only the latent itself - if a downstream node needs those values directly, that's another reason to reach for Select Latent Size Plus instead. And a quiet trap: these buckets are meaningful because of SDXL's training resolution - on a model with a different native size (SD 1.5's 512x512, or the 1–2MP range several 2026-era models run at), a "1MP" size here isn't the safe choice it is for Pony or Illustrious; check what your checkpoint actually wants before assuming this list applies. Beyond that, the usual batch caution: batch_size multiplies VRAM and time, it isn't a quality dial.