简单尺寸设置

Every txt2img workflow needs a width and a height fed into an Empty Latent, and if you've run more than a handful of shared workflows you know the pain: a dozen different "size" widgets, each with its own idea of what a good preset is, and you re-typing 1024 twice every time. EasySizeSimpleSetting is a two-INT answer to that. It's a single node whose entire job is to output width and height - and to give you three clean ways to decide what those numbers are.

The name is mostly honest: the whole pack ships exactly one node, displayed as "简单尺寸设置" in the UI. There are no models to download, no API to call, no key, no requirements.txt. It's pure plumbing, MIT-licensed, and small enough that it's worth reading the source before you trust it - the whole thing is one Python file and one JavaScript file.

What it does

A size router. You pick a size once, in one place, and wire the two INT outputs into anything that accepts a width/height pair - the stock Empty Latent, a third-party latent node, a ControlNet preprocessor, whatever. The README is explicit about what it is not: it doesn't create latents, resize images, crop, touch masks, detect models, or optimize VRAM. Two integers out. That's the deal, and it's the reason it can't break much.

The reason you'd actually reach for it over just typing numbers is Image mode. Feed it any IMAGE and it reads the tensor's original dimensions - image.shape[1] is height, [2] is width - and outputs them. For img2img that means your output latent matches your source image exactly, no math, no guessing, no accidental 1024-on-a-768-image mismatches. It reads the size and does nothing else; it never modifies the input image.

The three modes

• Preset - the headline feature. Sizes are organized as Family → Ratio → Resolution, with families for Common, SD1.5, SDXL, Pony, Illustrious, NoobAI, FLUX, FLUX.2, Krea, Krea2, Z-Image, Qwen Image, WAN, and Video. Pick a family, it filters the Ratio list, which filters the Resolution list - 67 resolutions at last count, all sensible, all multiples of 8.
• Image - inherits the source image's dimensions, as above.
• Custom - two plain INT widgets (1–8192) that pass straight through.

The frontend hides whatever doesn't apply to the current mode: pick Custom and the Family/Ratio/Resolution dropdowns vanish; pick Image and a lazy IMAGE socket appears. The image input only gets pulled when you're in Image mode, so it's free to leave unwired otherwise.

Where people get burned

The presets are all multiples of 8, but Custom mode doesn't check that. Type 1000 and you'll sail through the node and crash at the VAE/sampler with a dimension error. Same trap in Image mode: an odd-sized source image passes through untouched, and the failure just moves downstream. Keep your numbers divisible by 8.

Also remember Family is a catalog, not a model detector - the README says it plainly. Choosing "SDXL" doesn't load SDXL or validate against your checkpoint; it just offers SDXL-typical sizes. You can feed 1536×1536 to an SD1.5 model and it will happily try. That's on you.

Install

ComfyUI Manager → Custom Nodes Manager → search ComfyUI_EasySize → install → restart. If Manager can't find it - likely, since it's new enough to have zero community footprint - clone it manually:

cd ComfyUI/custom_nodes
git clone https://github.com/hadesybil-hub/ComfyUI_EasySize.git

Then restart ComfyUI. That's it: no dependencies, no models, no VRAM cost.

The honest take

This thing has essentially no community signal - one "Show and Tell" self-post, no replies - so if you're here, you almost certainly loaded a workflow that references it. That's fine: it's a featherweight, single-purpose utility that does one job cleanly, and the Image-mode inheritance is genuinely nicer than any big suite's size widget. It won't change your life, but it might stop you typing 1024×1024 for the rest of the afternoon.