PSNR, SSIM and LPIPS
Research papers score upscalers with PSNR (pixel error), SSIM (structural similarity, Wang et al. 2004) [1] and LPIPS (a learned perceptual distance, Zhang et al. 2018) [2]. LPIPS tracks human judgement better than PSNR or SSIM, which its authors call “simple, shallow functions” [2]. Vendor marketing rarely publishes any of these — so test yourself.
Every upscaler picks a side
Blau and Michaeli proved that no method can be both maximally faithful and maximally natural-looking [3]. GAN upscalers such as ESRGAN [4] and diffusion upscalers such as SUPIR [5] lean towards natural-looking detail; conservative models lean towards fidelity. Tools with both kinds — or a creativity dial — let you choose per image.
A fair five-image test
Run the same set through every tool you shortlist and compare at 100% zoom:
- a face (does the person still look like themselves?)
- small text (are the letters the same?)
- a fine texture — hair, fur, fabric
- a heavily compressed screenshot
- a line-art or anime image
