Pick the engine in the Model menu: Universal Upscale renders any factor from 1x up to the tested maximum 9.2x (or an exact target resolution), Universal Upscale Distilled is the 1.3 M parameter student that renders 9x natively in one pass on the same device as the default engine and resamples its output for other factors. Tiled inference keeps high-resolution upscaling inside memory (outputs are capped at 4096 px on the long side); the status line reports which model ran and what it did.