Aspect Ratios: Which Resolution Should You Use?

Why buckets matter

Stable Diffusion XL was trained on a small set of specific resolutions. Generate at 1024×1024 or 1344×768 and you'll get sharp, well-composed images. Generate at 1000×750 and you'll get soft, weirdly proportioned mush — because the model never saw a training image at that size.

The SDXL cheat sheet

Imagoat's Generate page uses these exact numbers behind the scenes — pick "16:9" and you get 1344×768, not some in-between size.

Upscaling later

If you need larger — a 4K wallpaper or a print-ready poster — generate at the correct bucket first, then run the result through an upscaler. Upscaling a well-composed 1024 image is dramatically better than generating at 2048 directly.

Portrait tip

For faces specifically, generate in 3:4 at 896×1152. It's the ratio the model saw most human portraits at during training, and the difference in facial coherence is huge.