What goes wrong when you ask for a picture much larger than the model's native size.
One pass past native repeats the subject across the frame
Past the size it was trained at, the model has no sense of what a whole body or a whole composition looks like at that scale. It repeats what it does know instead, so you get a second head or an extra arm, and skylines that repeat themselves across the frame.That is why a picture that is wrong at 1536 is often exactly right at 1024. Stable Diffusion 1.5 models start showing it past about 768 on a side, and the panel says so on the size control. SDXL holds on further and then breaks the same way.There are two ways out, and they are different tools:
The hires pass generates at a size the model is competent at, then enlarges that image and repaints it, so the composition is decided where it can be decided well.
An upscaler runs afterwards on the finished picture, which makes it bigger without making it different.
Use one of these whenever you want a picture larger than the model's native size.