How image upscaling works: interpolation, Lanczos and 'AI enhance'

Enlarging a photo means asking for pixels the camera never captured. Every upscaler answers that request differently — the classic ones by interpolating between the pixels you have, the new ones by guessing what should be there — and knowing the difference tells you when "enhance" is real and when it's fiction. The image upscaler uses the best of the classic methods; this guide explains why, and when to reach for the other kind.

The problem: pixels that don't exist

Doubling an image's width and height means each original pixel must become four. Where do their values come from? The information isn't in the file; a 2× enlargement contains exactly as much detail as the original, spread thinner. The only question is how to spread it so the result looks smooth rather than blocky, sharp rather than blurry — and no spreading method adds detail. The limits are physics, and What DPI you actually need for printing works through what that means for print sizes.

Interpolation methods, from blocky to sharp

• Nearest neighbour copies the closest original pixel: blocky, instant, and exactly right for pixel art, where blocks are the point — the pixel art scaler uses it deliberately.
• Bilinear averages the four nearest pixels: smooth and soft; the browser default for scaling images, and the reason a plain resize looks blurry.
• Bicubic fits a cubic curve through 16 neighbours: sharper edges, slight overshoot; Photoshop's long-time default.
• Lanczos uses a windowed sinc function over 36 neighbours (Lanczos-3): the sharpest of the classical methods, best at preserving edges without ringing, the choice for print work and what this site uses. Slower, but a second or two in a browser.

An unsharp mask afterwards counters the residual softness any interpolation adds. Together, Lanczos plus sharpening is what "high-quality resize" has meant for thirty years.

What AI upscalers actually do

Super-resolution networks (ESRGAN, Real-ESRGAN, the models behind phone "enhance" features and paid upscaling sites) are trained on pairs of small and large images and learn what detail usually looks like: the texture of skin, brick, grass, hair. Given a small image they hallucinate plausible high-frequency detail consistent with it. On the kinds of content they were trained on the results are remarkable; on faces they can produce a person who isn't quite the person, on text a word that isn't the word, and on anything unusual, confident nonsense. The detail is invented, which is a feature for a wallpaper and a problem for evidence, medical images or a photo of someone you love. They also need a large model and, usually, a server or a capable GPU.

Which method for which image

• Logos, icons, diagrams, screenshots: Lanczos, or better, retrace to a vector with PNG to SVG and scale infinitely.
• Scanned documents: Lanczos then a threshold; OCR benefits from the upscale.
• Photos for print at modest enlargement (up to 2×): Lanczos with sharpening — faithful and predictable.
• Photos for large, decorative enlargement: AI, accepting invented texture.
• Faces, evidence, anything that must be true: classical only, and be honest about the softness.

The honest limits of every method

A 200-pixel thumbnail will never become a 4K photo; a licence plate that is eight pixels wide cannot be read by any algorithm, whatever television says. The realistic gains are 2× with no visible loss on a good source, 4× for graphics and scans, and beyond that either softness (classical) or fiction (AI). The best upscale is the one you don't need: keep originals at full resolution, and compress copies for the web rather than shrinking the master.