Empty Latent Size Picker

Every ComfyUI beginner learns the same chore: drop in an Empty Latent Image, type the width, type the height, and hope you didn't mix the two up. EmptyLatentSizePicker replaces that whole three-node dance with one node. You pick a resolution from a dropdown, it hands you a ready-made blank latent and the matching width and height as separate integers.

It's the extended version of the pack's Simple Size Picker - same 34-preset dropdown, but the output is a real LATENT tensor rather than just two numbers. If you're generating anything, this is the one you actually want, because it plugs straight into a sampler. No EmptyLatentImage, no math nodes, no ratio fumbling.

Mechanically it's straightforward. The node builds a tensor of zeros shaped [batch, 4, height // 8, width // 8] - that's the standard 8× latent compression every VAE-family model uses, so a 1024×1024 selection becomes a 128×128 latent with four channels. It's created on ComfyUI's chosen device and handed to you as a normal latent dict, exactly what a KSampler expects. A zero latent means "start from scratch," which is what you want for text-to-image.

The inputs that matter:

• resolution - the same curated dropdown as Simple Size Picker: 34 entries from 512 x 512 up to 9600 x 1200, grouped by aspect ratio.
• batch_size - defaults to 1, and the node will happily go up to 4096. Please don't actually try 4096 on consumer hardware; you'll melt your VRAM before the sampler gets a turn.

Outputs: LATENT (the blank latent for your sampler), width, and height (both INT, handy if a downstream node wants the numbers for conditioning or upscaling math).

Install it via Custom Nodes Manager (search ComfyUI_BishaNodes) or clone it into custom_nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/vladpro3/ComfyUI_BishaNodes.git

Then restart. No model files, no extra dependencies, MIT-licensed - it's a self-contained utility.

The honest caveats: a blank latent is still a blank latent. Picking 3840 x 2160 here does not mean SDXL can generate 4K cleanly - the dropdown keeps you honest about ratios, but you still have to respect your model's native resolution and upscale afterward if you want the big canvas. Flux users should also note a couple of the presets (like 1080 x 1920) aren't divisible by 64, which Flux cares about more than SDXL does. And if you're on a model with a 16-channel VAE, the hardcoded 4 channels won't match - but that's an edge case well outside this node's target audience. For the 99% case of "generate at a sane SDXL/Flux-ish size," this node is the cleanest one-stop size solution in the pack.