Omini Kontext Image Scale

There's not much to this node, and that's the point. The Omini Kontext Image Scale takes any image and rescales it to the resolution bucket closest to its aspect ratio - one input, one output, no settings, done. It exists because the omini-kontext pipeline works best when your scene and reference images land on the resolutions the model actually saw in training, instead of whatever arbitrary pixel dimensions your source image happens to be.

How it works

The node keeps a hardcoded list of seventeen "Kontext resolutions", all on the 64-multiple grid the model likes - it runs from portrait (672×1568) through square (1024×1024) to landscape (1568×672). It computes your image's aspect ratio, finds the bucket whose width:height ratio is closest, and resizes to it using ComfyUI's built-in upscaler in lanczos mode. That's the whole mechanism. No sharpening, no face restore, no tiling.

The one detail worth knowing: it will happily upscale a small image as well as downscale a big one. If your character reference is 512×512, this node quietly stretches it to a bucket and the model doesn't care - the reference gets resized again inside the pipeline anyway. So think of this as a "put the image on the grid" utility rather than an upscaler.

Where it slots in

Wire it between a Load Image and the main Omini Kontext Pipeline node, on either your scene or your reference. The bundled example workflow puts it after one of the image loads, which tells you it's optional prep rather than a requirement - the pipeline will handle arbitrary sizes on its own, just less predictably. The README lists it under "Prepare Images" in the basic workflow: load scene, load reference, optionally scale both.

Honest verdict

If you're already feeding 1024×1024 square images from the same generator, this node is doing nothing for you. It earns its keep when your source material is an odd phone-picture or a downloaded asset with a weird aspect ratio - snapping it to the grid before encoding removes one variable when you're debugging why a character insertion came out wonky. It's also the safest thing in this pack to run on a low-VRAM machine, since it does nothing but tensor math. Zero risk, small benefit, two seconds of setup. That's the whole review.

Installing the pack is the standard dance - clone into custom_nodes, pip install -r requirements.txt, restart. If you came here from the other Omini Kontext nodes you already have it.