LZ Preset Empty Latent Image

Every generation starts with an empty latent - a tensor of pure noise-shaped zeros that the sampler progressively denoises. ComfyUI's built-in EmptyLatentImage does this fine. So why would you install a pack for a blank canvas? Because the built-in one makes you remember the math, and this one does the remembering for you. Preset Empty Latent Image is the same core node with a dropdown of nine resolutions, all deliberately close in total pixel count to 1024x1024 - which is the entire trick.

The reasoning is straight out of how these models are trained. Each architecture has a native resolution, and going far off it is where artifacts start showing up - double heads, stretched bodies, repeated patterns. SDXL and the models that followed were trained around a one-megapixel band, so what actually matters is keeping total pixels near 1MP while swapping the aspect ratio to suit your subject. A portrait shot wants 896x1152 (7:9); a landscape wants 1152x896 (9:7) or wider at 1344x768 (7:4). This node pre-bakes those ratios so you're always within spitting distance of the pixel count the model expects, and every preset is a multiple of 64, which keeps the latent cleanly divisible by the 8x spatial downscale.

Mechanically it's the stock recipe: torch.zeros of shape [batch, 4, height//8, width//8] - four channels, the classic VAE channel count, placed on ComfyUI's intermediate device - returned as a LATENT with the downscale_ratio_spacial flag so downstream nodes know the spatial compression. It's not magic, and it knows it.

The inputs you actually touch:

• resolution - pick from 1024x1024 (1:1) up through the 12:5 ultrawides. That's the whole reason the node exists.
• batch_size - default 1. The slider goes to 4096; unless you're deliberately generating a big batch, don't. The model multiplies it, and VRAM disappears fast.

One LATENT output, straight into a sampler.

Install is the pack's standard, trivial path - Manager search comfyui-uonr-nodes, or:

cd ComfyUI/custom_nodes
git clone https://github.com/uonr/comfyui-uonr-nodes

then restart ComfyUI. No models, no dependencies; the README calls it a personal node pack, and it shows in the best way.

The honest take: this is a convenience wrapper, not a capability. If you're comfortable typing 1216x832 from memory, the stock node is fine. If you'd rather click than type, or you're tired of pasting the same ratios into every workflow, this is the version you'll actually reach for. Just don't use it with SD3/Flux - those want the 16-channel sibling from the same pack, and feeding them a 4-channel latent errors before the first step.