Native resolution

The size a model was trained at, which is the size it knows a whole scene at.

Training happened at a fixed size, and a model's competence is built around it: around 512 a side for Stable Diffusion 1.5 and its descendants, and 1024 for SDXL.Read it as the scale at which the model knows what a whole scene looks like. It is neither a minimum nor a recommendation.The examples walk one SDXL request down from its native 1024, and each size gives a different picture:
  • at 512, bodies are still drawn correctly, but the framing tightens into a close-up of whatever fits;
  • at 256 you get a background with no recognisable subject in it;
  • at 128 the image is unreadable.
Going the other way breaks differently, by repetition, which is what covers.The panel knows both native squares and says so. It snaps an untouched size across when you switch to a model of the other family, and it notes when your dimensions have drifted out of the one you are on.A moderate change is fine, and every aspect preset is a moderate change. Halving or doubling a side is where the picture breaks down.
Size · Interactive explorer Size
The same picture rendered at 512x512Size 512x512 512 × 512
128x1281024x1024

Anatomy holds, but the framing collapses to a cramped close-up of whatever fits.

Fetching the example images.

In the glossary

Browse the model list