Nearest Qwen Resolution

A simple ComfyUI custom node that takes an IMAGE input and outputs the width and height (two INTs) of the nearest recommended Qwen-Image resolution, matching orientation and minimizing aspect ratio difference. Ties are resolved by smallest size delta to the input image.

• Input: IMAGE (tensor shaped (B,H,W,C); first frame used)
• Output: width (INT), height (INT)
• Setting: square_tolerance (percent; how close to 1:1 counts as square)

How it decides

1. Determines orientation (Wide / Vertical / Square) from the input size with a configurable tolerance for square.
2. Filters the Qwen presets to that orientation.
3. Chooses the preset with the smallest aspect ratio difference; ties broken by the smallest difference to the input dimensions.

Presets

These resolutions are the official recommended sizes for Qwen-Image models, specifically optimized during the final training phase. They cover common aspect ratios like 9:16 (928x1664), 1:1 (1328x1328), and 16:9 (1664x928), ensuring stable generation quality without artifacts.

Why these sizes?

• Based on 32-pixel multiples (~1-2 megapixels), allowing efficient processing by diffusion models
• Other common social media formats (e.g., 1080x1080 or 1080x1350) often produce inferior results, as they aren't aligned with Qwen's training data

Available presets:

• Vertical: 928x1664 (9:16), 1056x1584 (2:3), 1140x1472 (3:4)
• Square: 1328x1328 (1:1)
• Wide: 1664x928 (16:9), 1584x1056 (3:2), 1472x1140 (4:3)

Category

Appears under image/resolution.

Notes

• Qwen-Image does not require 64-multiple rounding, so widths/heights are reported exactly as the preset.
• If the input is very close to square (within square_tolerance%), the square preset is preferred.

License

MIT License - see LICENSE for details.