🔶 Empty Latent Image - Video Size

The stock Empty Latent Image node makes you type raw pixel numbers into two boxes and hope you picked something your model actually likes. This one flips that around: you pick an aspect ratio, then pick a size tier, and it works out the width and height for you. It's a small thing, but it's the kind of small thing that saves you from accidentally generating at 1920x1080 on an SD 1.5 checkpoint and wondering why you got a duplicated torso.

What it does

Two dropdowns. Mode picks the aspect ratio: Widescreen (16:9), Portrait/Smartphone (9:16), or Width = Height (square). Size picks how big: 360p, 480p, HD, Full HD, QHD, 4k, or 8k. Combine the two and it computes the matching width and height for you, outputs a blank LATENT at that size, and also hands back WIDTH and HEIGHT as raw INTs so you can wire the same numbers into an upscaler or a resize node further down the graph.

Here's the part worth reading before you pick a tier: Full HD, QHD, 4k, and 8k are all flagged in the dropdown itself as "Check Info" - and 8k specifically says "Latent HD." That's the node telling you it isn't generating an actual 4k or 8k latent, because no SD 1.5 or SDXL checkpoint produces a coherent image at those sizes in one pass - you'd get the tiling and duplicated-anatomy mess any diffusion model gives you well past its native resolution. Instead, past a certain tier the node keeps the LATENT output at a workable generation size and reports the size you actually asked for through WIDTH, HEIGHT, and the Info string, so you can generate small and then feed those numbers into an upscale step to hit your real target. It's the same "generate native, upscale after" pattern that's standard advice everywhere else in ComfyUI - this node just bakes it in instead of making you remember it.

Why you'd reach for it

Mostly convenience and mostly for people still on SD 1.5/SDXL-era workflows, where native resolution actually matters and going off it produces visible artifacts. If you already know your model's trained resolution and are comfortable typing 512x512 or 1024x1024 into the stock node, you don't strictly need this. Where it earns its keep is exactly the animation-adjacent workflows this pack is built for - pick "Widescreen / HD" once, wire WIDTH and HEIGHT into whatever else in your graph needs matching dimensions, and you're not maintaining the same two numbers in three places.

Installing it

Via ComfyUI Manager: search "Chaosaiart-Nodes," or use "Install via Git URL" with https://github.com/chaosaiart/Chaosaiart-Nodes, then restart. Manually: cd ComfyUI/custom_nodes && git clone https://github.com/chaosaiart/Chaosaiart-Nodes and restart. On Linux, run pip install opencv-python and pip install tqdm (with your venv active if you have one) - not needed by this node specifically, but other nodes in the same pack import them, so skipping this step will throw errors when ComfyUI loads the pack. Windows has a bundled Install_windows script. No models to download for this one.

Troubleshooting

If your output looks stretched, tiled, or duplicated at the higher size tiers, that's not this node malfunctioning - it's the underlying checkpoint being pushed past its native resolution, which is exactly what the "Check Info" tiers are trying to warn you away from. Stick to 360p/480p/HD for direct generation on an SD 1.5 checkpoint, or SDXL's native 1024-class sizes, and upscale afterward for anything bigger. And remember this node only outputs a blank latent - it doesn't do the upscaling itself, so you'll still need a separate upscale node (ESRGAN, hires-fix-style two-pass, or ControlNet Tile) if you want to actually land at the Full HD/4k/8k number it reported.