Upscaling AI Images — Which Method for Which Job

Three families, three strengths

Pure-pixel upscalers (ESRGAN, RealESRGAN, Swin2SR) — fast, cheap, preserve the image faithfully. Best for logos, illustrations, and images where you must not invent detail.

Diffusion-based tile upscalers (Ultimate SD Upscale, MultiDiffusion) — slower, more expensive, invent plausible detail. Best for photos where a little hallucinated texture is fine.

SUPIR — the current state of the art. Very slow, very expensive, near-photographic detail restoration from tiny sources. Overkill unless you actually need 4K.

Which to pick

The 2× rule

Upscaling more than 4× in a single pass is a bad idea, no matter which method. Chain two 2× passes instead — cleaner result, fewer weird artifacts.

When to just generate larger

If you know you need a 2048×2048 final: sometimes it's better to generate at 1024, upscale to 2048 via Ultimate SD Upscale with a low denoise. If you generate at 2048 natively in SDXL you get double heads and other resolution drift bugs (SDXL was trained at 1024).

Flux handles larger native resolutions better but costs more per image — do the math.