Qwen Image Natives Resolutions

Diffusion models have native resolutions: the sizes they were actually trained on, where composition, text rendering, and coherence all come out best. Stray far from them and Qwen Image starts doing what models do when you ask for the impossible - garbage. This node hands you the exact list of sizes Qwen Image was trained on, as a dropdown, and outputs the one you pick as a packaged resolution.

The name is a mouthful, but it's honest about the job. IG1QwenImageNativesResolutions generates nothing, calls no API, needs no key. It's a lookup table with a resolution-shaped output, which is genuinely useful because nobody memorizes this list - and the list isn't what you'd guess.

Where the list comes from

The author lifted it straight from the Qwen team's own answer in the QwenLM/Qwen-Image GitHub issues - the source file even links the comment - and hardcoded the seven training resolutions. Sorted biggest-first, the dropdown reads roughly like:

• 1328×1328 (1:1 @ 1.76MP) - the default, and the community's de facto standard. Reddit comparisons routinely generate Qwen at 1328×1328 specifically because it's the native square.
• 1584×1056 (3:2 @ 1.67MP) and its portrait twin 1056×1584
• 1472×1104 (4:3 @ 1.63MP) and 1104×1472
• the wide-ish 1664×928 and 928×1664, the pair people forget exist

Notice there's no 16:9 and no clean 2:1. Qwen's training resolution set just isn't built around your monitor, which is exactly why a lookup node earns its keep here.

What it does

The one input, native_resolution, is that dropdown, defaulting to 1328×1328. The one output is a single RESOLUTION value - a custom IG1_RESOLUTION type bundling width and height into one wire. On its own that's not much to look at, but the pack's other nodes make it sing: feed it into Resolution Properties to unpack WIDTH, HEIGHT, MEGAPIXELS, and ASPECT_RATIO for your EmptyLatentImage or conditioning, or hand it to the Resolution Advisor if you want a HiRes plan built around your chosen native size.

Worth knowing about the model underneath: Qwen Image is Alibaba's 20B Apache-2.0 flagship from the Qwen/Tongyi team - the family whose Edit line became the open editing standard and, being Apache, the one you can fine-tune without license anxiety. That's the reason it became the 2025–26 alternative to Flux for people who wanted open weights without BFL's terms.

Installing it

Standard pack story: ComfyUI Manager → Install via Git URL → https://github.com/iguanesolutions/comfyui-flux-resolution.git, or:

cd ComfyUI/custom_nodes
git clone https://github.com/iguanesolutions/comfyui-flux-resolution

then restart. No model files, no weights - this node is pure Python constants. But that means it doesn't give you Qwen Image itself; you still need the actual checkpoint loaded in ComfyUI to generate. This node just stops you from asking it to work at a resolution it was never trained at.

Small print

The list is frozen as of the pack's version. If the Qwen team ever retrains at new sizes, this node won't know until the author updates. Given Qwen-Image-2.0 is API-only as of this writing, the v1 training resolutions should stay accurate for a while - which is about as stable as anything gets in this ecosystem.