Qwen Aspect Ratio (Creepybits)

Pick "16:9" on most aspect-ratio nodes and you get 1920×1080. Qwen Aspect Ratio picks 1664×928. That difference is the entire reason this node exists.

Qwen-Image is trained to produce its best output at a specific, non-standard set of resolutions, and the Qwen-Image entity in the KB confirms the model family's quirks are real. Generating at those exact dimensions avoids artifacts and weird compositions. This node is a preset manager for precisely those values: you choose a familiar aspect ratio from a dropdown, and it outputs the exact Qwen-native width and height as integers, ready to feed your Empty Latent (or the sampler's width/height fields).

How it works

It's a static lookup table - no computation, no model, no API. Seven ratios, each mapped to the resolution Qwen-Image expects:

• 1:1 → 1328×1328
• 16:9 → 1664×928
• 9:16 → 928×1664
• 4:3 → 1472×1104
• 3:4 → 1104×1472
• 3:2 → 1584×1056
• 2:3 → 1056×1584

ratio is a dropdown of those seven options. Outputs are width and height as INTs. Wire them into any width/height input - most commonly the Empty Latent Image node or your KSampler's resolution settings. The node itself is instant and can't fail; it's a glorified constants table with a nice UI.

When to use it - and when not to

Use it whenever you're generating with Qwen-Image or Qwen-Image-based stacks. Those native resolutions are what the model was tuned on, and they're deliberately not "clean" numbers like 1024 - that's exactly the kind of thing you don't want to memorize or round.

That said, these are just integers when they come out. Nothing stops you from using 1328×1328 on an SDXL graph if you want that aspect - the node doesn't enforce anything, it just tells you the numbers. If you're on a model with its own preferred resolutions (SDXL, Flux), the more general habit is to use that model's native sizes instead; this node's value is specifically knowing Qwen's.

One practical note: 1664×928 and friends are larger than the 1024-class sizes most people default to, so watch VRAM if you're on a small card - the node doesn't resize anything, it just hands you dimensions your sampler has to honor.

Installing it

Part of ComfyUI-Creepy_nodes:

• ComfyUI Manager: search "Creepy_nodes" and install.
• Manual:

  cd ComfyUI/custom_nodes
  git clone https://github.com/Creepybits/ComfyUI-Creepy_nodes.git
  

Restart ComfyUI. Zero dependencies, zero setup, works offline.

Troubleshooting

There's almost nothing to break. If width/height aren't reaching your Empty Latent, check that you wired both outputs (they're separate sockets). And if your Qwen generation looks off despite using these values, remember the model also cares about its prompt format - that's where the pack's QWEN Prompter node is meant to come in. The aspect ratio is one half of the recipe.