ICHIS Aspect Ratio Plus

Every SDXL workflow starts the same way: an Empty Latent node, and you frantically trying to remember which pixel dimensions are actually 3:4. The ICHIS Aspect Ratio Plus node from the Ichis pack is that node with the cheat sheet welded on. Pick a ratio and it hands you the correct width, height, and a ready-to-sample latent - and if you want, it will roll through ratios for you across batches instead of you editing the widget by hand every time.

What it actually does

It's a drop-in replacement for core EmptyLatentImage, except the aspect_ratio dropdown carries the eight SDXL-recommended presets (1:1 square 1024x1024, 3:4 portrait 896x1152, 9:16 portrait 768x1344, 16:9 landscape 1344x768, 21:9 landscape 1536x640, and the rest). Pick one and it outputs:

• WIDTH and HEIGHT - the pixel dimensions, handy if you're doing size math downstream.
• LATENT - a blank 4-channel tensor at 1/8 resolution. SDXL's VAE downsamples 8x per side into 4 latent channels, so the node builds it the exact same way core ComfyUI does, then KSampler adds the noise. Feed it straight into the sampler.
• SELECTED_RATIO - the human-readable name of what won.
• NEXT_STEP - which step index is coming up next, if you care to display it.

The inputs that matter

• aspect_ratio - the preset picker. This is the one you'll set 95% of the time.
• upscale - a plain multiplier on both dimensions (0.1–10). One catch: the result gets integer-truncated and the latent is width // 8 on each side. Upscale to something not divisible by 8 and your reported WIDTH won't quite match what decodes. Keep multiples of 8 unless you like off-by-one geometry.
• mode - normal (use the picker), step (cycle through the enabled ratios one per queue, wrapping around), or random (seeded choice).
• size_mode and the include_* toggles - filter which ratios are eligible when cycling or rolling randomly. You can, for example, restrict to landscape only and step through them.
• seed (random mode), reset_step (jump the step counter back to zero), batch_size (how many identical latents).

One behavior worth knowing: in normal mode the node caches like any ordinary node, but step and random force themselves to re-run on every queue because their IS_CHANGED returns a fresh timestamp. That's deliberate - a stepped or random ratio has to execute even if nothing else changed. If everything seems to re-run every time, check you haven't left it in random.

How to install it

Install the whole pack - all the ICHIS nodes ship together:

cd ComfyUI/custom_nodes
git clone https://github.com/ichabodcole/ComfyUI-Ichis-Pack

Then restart ComfyUI. You can also just search for ComfyUI-Ichis-Pack (or "Ichis") in ComfyUI Manager and click install. There are no model downloads and no heavy dependencies - the pack's only real dependency is torch, which you already have.

Troubleshooting

The latent is zeros, not noise - that's correct and matches core EmptyLatentImage; KSampler adds the noise. If your images come out a different size than the picker promised, you upscaled to a non-multiple-of-8 dimension. And if random mode ignores your picks, remember the include_* toggles and size_mode filter the pool; with everything toggled off it silently falls back to the manually selected ratio.