Random SDXL Latent

Every SDXL user has the nine native aspect ratios memorized: 1024×1024, 1152×896, 1216×832, 1344×768, 1536×640, and each one's rotation. The built-in Empty Latent Image node makes you type those numbers by hand every single time. Random SDXL Latent does the opposite - you give it nothing, it rolls the dice across all nine sizes, and hands you a ready-to-sample latent plus the width and height it chose.

That's the whole pitch: an easy way to randomize canvas shape across a batch. Fixed seed, same prompt, nine different crops of the image and you get a surprisingly fun contact sheet. If you're the kind of person who runs a batch and eyeballs the winners, randomizing resolution is a cheap extra axis of variation that costs nothing but time.

How it works

The mechanism is almost embarrassingly simple, and that's a feature. It picks a random (width, height) pair from the nine SDXL-native dimensions, then builds a pure Gaussian noise tensor at the VAE's latent scale: torch.randn(1, 4, height/8, width/8). The /8 is because the SDXL VAE downsamples 8× per side - a 1024×1024 image lives in a 128×128, 4-channel latent. That tensor is exactly what a KSampler expects as its latent_image input. The node also flags itself as always-changed, so it rerolls every queue run instead of handing back a cached size.

The inputs and outputs

There are no inputs. Not a resolution field, not a seed, not a batch count - the "random" is the entire feature. You get three outputs:

• LATENT - wire this into the latent_image socket of a KSampler.
• WIDTH and HEIGHT - the pixel dimensions of the roll, so you can display them in the graph or use them to drive other text nodes.

Gotchas

• No seed, no reproducibility. The latent shape rolls each run. Your KSampler's own seed still controls the actual image content - this node only randomizes the canvas - but if you need the same aspect ratio back, don't look for a way to lock it; there isn't one.
• It's SDXL-only, by construction. Those nine dimensions are SDXL's trained ratios. If your checkpoint is SD 1.5-class (native 512) or a newer model with different expectations, feeding it a 1536×640 latent will give you stretched or broken results. This is an SDXL workflow tool.
• Batch size is fixed at 1. No batch knob, so if you want a batch of latents you'll be rolling one per image anyway.

Install via ComfyUI Manager (search ComfyUI-Bricolage) or cd ComfyUI/custom_nodes && git clone https://github.com/sheldongriffin/ComfyUI-Bricolage, then restart. No model downloads. Given that core ComfyUI already has Empty Latent Image with the same nine values one dropdown away, this node is a nicety rather than a necessity - but for batch experiments it removes the fiddliest part of the workflow, which is exactly the kind of small convenience the pack's author says these nodes are for.