CNP Random Resolution

CNP Random Resolution picks a width and height for you, on a grid, from a seed you control. The point is exploration: instead of hand-setting 1024×1024 and hoping, you let a run grab a random aspect ratio from a sensible list - 512×1024, 1024×1536, 2048×1024 - and see what the model does with it. That's the batch-variation workflow ComfyUI is uniquely good at, except here the thing varying is the canvas, not just the seed.

Everything it emits is a multiple of base_unit, so with the default 512 you get 512, 1024, 1536, 2048 - all multiples of 64, which keeps the VAE happy. No off-grid sizes that make latents or the VAE complain.

Three modes

• fixed - uses fixed_width_steps and fixed_height_steps directly. No randomness at all; effectively a nicer way to type a resolution.
• random - picks a step count for width and for height independently.
• random_ratio (default) - builds a curated set of aspect ratios from common pairs (1:2, 2:3, 3:4, and their landscapes) then picks one. This is the "surprise me, but make it sane" mode.

Inputs that matter

• mode - the three above.
• base_unit - the grid size. 512 for SD1.5-era models, 1024 for SDXL/Flux native resolution. Step is 64, so you can't emit a dimension the VAE will choke on.
• min_steps / max_steps - the range of base_units per side. Defaults 1–4, meaning 512–2048 at base 512.
• orientation - any, square, portrait, or landscape. Force the shape when you need it.
• ratio_mode - common (both orientations), vertical_common, horizontal_common, or free (any combination of steps).
• seed - same seed, same resolution, every time.

Outputs

width (INT) and height (INT) go straight into the Empty Latent Image node's width/height inputs. There's also a resolution STRING ("1024 x 1536"), handy for a caption or to hand to CNP Format when you're saving images named by size.

The gotcha that will trip you up

The seed is a plain INT widget, not a ComfyUI seed control. There's no "randomize after generate" built in, so with the default seed of 0 you'll get the exact same resolution on every run. Great for reproducibility, very confusing if you expected a die roll. To actually randomize, feed it a fresh seed each run (bump the value, or wire in a random-seed node) - that's the whole trick to making this node do its job.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/chemicalshock/ComfyNodesPlus

or ComfyUI Manager → "ComfyNodesPlus", then restart. No requirements.txt and no model files - it's pure Python stdlib, so it can't break your environment like the heavier utility suites.

Troubleshooting

• The node has no idea how much VRAM you have. base_unit 2048 with max_steps 4 is 8192² - a 67-megapixel latent that will OOM most cards. Dial max_steps down.
• min_steps greater than max_steps? It silently swaps them rather than erroring.
• It emits one resolution per node execution, and a fixed sampler batch shares it. For per-image variety in a batch, randomize the seed between runs rather than expecting one node call to vary mid-batch.