Bori Qwen Image Resolution

Qwen-Image is Alibaba's 20B diffusion transformer - the family famous for text rendering, the Edit line that became the open editing standard, and Apache 2.0 weights you can actually fine-tune. It's also, like every diffusion model, trained on a fixed set of native resolutions. Stray far outside them and you get the usual tells: duplicated anatomy, extra limbs, tiling. The buckets Qwen actually recommends are a specific seven, and they're not round numbers. 1328x1328. 1664x928. 1472x1140. Hand-typing those into an Empty Latent Image node on every run is exactly how you end up with a typo'd 1624x928 and a slightly cursed image you can't quite explain.

That's the entire job of Bori Qwen Image Resolution: one dropdown, seven correct values, two INT outputs. It's the ComfyUI equivalent of a sticky note that already has the right numbers on it.

What it actually does

The node is gloriously simple, and "gloriously" is doing real work here. There's no API, no model files, no keys, and no Python dependencies - the whole thing is a hardcoded lookup table. The seven entries from Qwen's supported-resolution list (each stored as width, height, and aspect ratio) show up as a dropdown. You pick one, the node splits the string on the "x", and returns width and height as integers. That's the entire implementation: a string split on a dropdown value, and it's the right amount of code for the job.

The seven resolutions, for the record

| Choice | Aspect | Pixels | |---|---|---| | 1328x1328 | 1:1 | 1.76 MP | | 1664x928 | ~16:9 | 1.54 MP | | 928x1664 | ~9:16 | 1.54 MP | | 1472x1140 | ~1.29:1 | 1.68 MP | | 1140x1472 | ~1:1.29 | 1.68 MP | | 1584x1056 | 3:2 | 1.67 MP | | 1056x1584 | 2:3 | 1.67 MP |

Notice they all cluster around 1.5–1.8 megapixels. That's Qwen-Image's native zone, and community testing lands on 1328x1328 (~1.77 MP) as the sweet spot for consistency. It's the same "generate at native resolution, upscale after" advice the community has been giving since SD 1.5 - just pre-computed for you, so you don't have to remember which of these odd numbers pairs with which aspect.

Inputs and outputs

One input, one dropdown: resolution, with the seven choices above. That's the whole input surface - no optional inputs, no hidden toggles.

Two outputs:

• width (INT)
• height (INT)

Wire both into the width/height slots of an Empty Latent Image node in your Qwen-Image workflow. Some Qwen loader wrappers expose their own width/height fields, and the same two ints feed those fine. If your setup has a conditioning prep node that wants a resolution string, the dropdown value itself is handier than the split ints - but for the standard latent path, width and height are exactly what you need.

Installing it

Dead simple, because there's nothing to install beyond the files:

cd ComfyUI/custom_nodes
git clone https://github.com/boricuapab/ComfyUI-Bori-QwenImageResolutions

Restart ComfyUI. That's it - no requirements.txt, no pip install, no model downloads, zero VRAM cost. ComfyUI Manager can find it too: search the pack title "ComfyUI-Bori-QwenImageResolutions" and hit Install. One nice thing about a node that's just a dropdown is that it can't drag a transformer version conflict into your environment with it.

Common issues, such as they are

This node is so thin that the failure modes are mostly "you're using it wrong":

• It does nothing for non-Qwen models. These numbers are Qwen-Image's recommendations, not universal law. If you're sampling Z-Image or Flux, 1664x928 is just a number that happens to be in a dropdown. Match the node to the model actually in your graph.
• It's not a free-form canvas. You get fixed 1:1, 16:9, 3:2 and in-between buckets, not any size you want. If you need an exact canvas, pick the closest bucket and crop or outpaint afterward - standard advice for every architecture.
• Manager can't find it. Try searching just "Bori" or "Qwen Image Resolution" rather than the full repo name. It's a small, low-profile pack, so the registry entry can lag.

The honest caveat: with effectively zero search traffic and a single seven-item dropdown, this is a tiny QoL pack from an obscure author, not a must-have. You could replicate it with a Primitive node and a text box in two minutes. What the dropdown buys you is discoverability and the absence of typos - and for a value you're going to enter into every single Qwen-Image run, that's a real, if modest, win.