Bicubic vs neural
Bicubic interpolation fills new pixels by averaging surrounding ones. It's fast and produces smooth, slightly soft enlargements. Neural upscaling (Real-ESRGAN, ESRGAN, SwinIR) hallucinates detail based on a learned prior — the results look sharper but the detail isn't real.
When upscaling helps
- Old photos for print — neural fills in detail bicubic can't
- Small assets for high-DPI displays — 2× usually enough
- Recovery of overcompressed JPGs — neural removes blocking
When it doesn't
Faces and text are the failure modes. Real-ESRGAN smooths skin too aggressively (everyone looks 23) and turns small text into convincing-looking gibberish.