HiDream Resolution

Where Flux is a megapixel model that'll take almost any dimension you hand it, HiDream is the opposite - it has a set of hard-coded resolutions it was trained on, more like SDXL. Wander off those and quality drops. This node is just a clean dropdown of the sizes HiDream actually likes, so you pick an aspect ratio and get the correct width and height back. No arithmetic, no guessing which "portrait" HiDream means.

If you're only running Flux, you'll never touch this one; it's here for the HiDream corner of the pack. But it's exactly the kind of quality-of-life node ControlAltAI builds - small, opinionated, saves you a lookup.

How it works

One input, resolution, which is a dropdown of seven named aspect ratios: 1:1 (Perfect Square), 3:4 (Standard Portrait), 2:3 (Classic Portrait), 9:16 (Widescreen Portrait), 4:3 (Standard Landscape), 3:2 (Classic Landscape), and 16:9 (Widescreen Landscape). Pick one and the node returns the exact dimensions HiDream expects for that shape.

That's the whole mechanism. The value is that these aren't arbitrary - they're the trained sizes, so you're not fighting the model by feeding it something it never saw.

The inputs and outputs that matter

• resolution - the only knob. Choose your aspect ratio.

Outputs: width (INT), height (INT), a resolution string, and a preview image that's just a visual swatch of the chosen ratio. Send width and height into your empty latent; ignore the preview unless you want the visual confirmation. Same output shape as the Flux Resolution Calculator, so if you've used that, this feels identical.

How to install it

ComfyUI Manager → search ControlAltAI Nodes → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/gseth/ControlAltAI-Nodes

then restart. Pure Python, no downloads. You will, of course, need HiDream itself set up separately for the resolutions to be useful - this node just picks numbers.

Common issues

There's not much to break in a dropdown that returns two integers. The thing worth saying is about HiDream as a model rather than this node: HiDream (the I1 line and the later O1) is heavy and has largely stayed a niche pick next to Z-Image and the Flux 2 family, so if you're setting this up in 2026, know you're on a road less traveled. HiDream also tends to want its native sizes strictly - which is the whole reason this node exists - so don't try to "improve" the output by manually bumping the dimensions past what the dropdown gives you. Feed it what it was trained on and let an upscale pass handle anything larger.