Image Upscaling: What Actually Improves and What Is Pure Imagination

Plain resampling computes new pixel values between existing ones — bicubic, lanczos and browser smooth-scaling all work this way. The result is larger, correctly proportioned, and soft, because a mathematical guess at the missing middle is exactly what it sounds like. Resampling is honest: it never pretends to know more than the file does.

AI upscaling trains a model on millions of photo pairs (small and large versions of the same scene) until it learns what a soft edge usually has on either side. Given a small image, it does not magnify — it predicts. A model that has seen a thousand eyelashes draws better ones than your 320 px photo contained. Sometimes the prediction is uncanny. Sometimes it invents a texture that was never there, and at 4× on faces the invention becomes visible as a porcelain sheen.

Knowing which tool answers your need is the whole skill: resampling for structure (graphics, screenshots, anything with straight edges and text), prediction for organic subjects (portraits, landscapes, fabrics) where plausible detail is indistinguishable from real.

• Can: smooth stair-stepped diagonal edges; turn blocky 8×8 JPEG mosaics into continuous tone; make a 500 px photo fill a 1500 px container without visible pixels; clean modest noise while enlarging.
• Cannot: read a licence plate that occupied 12 pixels; restore faces crushed into JPEG mush; invent facial identity the model has never seen (models can enhance and cannot reconstruct specific features); turn a 100 px icon into a billboard.
• Can guess: fabric weave, skin pores, foliage — organic patterns the training data covers richly. Treat these as a painting of the original, not a scan of it: fine for decoration, wrong for evidence.
• Cannot honestly be fixed by any upscaler: severe motion blur (directionally smeared information is gone), chromatic aberration at edges, and focus failure — a blurred subject is blurred because the lens never recorded the edge.

The tempting move is maxing the slider: 2× looks plausible, so 8× must look better. It does not. Every pass compounds the model's inventiveness, and quality is not linear — a 2× pass plus a careful resample to odd sizes beats a single 8× pass on anything that matters.

Practical defaults: one step (2×) for photos headed to print from a decent original; two cautious steps max for old family photos; and for logos and graphics, reach for vectorising instead of magnifying — a raster logo scaled past about 4× is telling you it wants to be an SVG.

One quiet win worth knowing: an image that is only slightly too small (15-25%) is nearly invisible at display size on the web, because screens are denser than the design era that set these habits. Confirm the real display width before upscaling anything — half of 'please enlarge' jobs are nobody's problem once measured.

Start from the largest clean original — not a screenshot, not a re-saved chat image, because those have already lost generations you cannot get back. Upscale once, modestly (2×), then inspect against the use: on screen at 100%, at the intended display width; for print, at the printer's native preview.

Order matters for the finishing steps: upscale → resize to the exact final dimensions (the extra scale gives you room for a good resample) → apply a light final sharpening if the result looks soft → compress for delivery. Sharpen first and you sharpen noise; compress first and the upscaler spends its prediction budget repairing blockiness.

Keep a copy of the untouched original — permanently. Upscaling is a one-way, lossy-ish operation on the pixels you fed it, and tomorrow's better model deserves today's best file.

The most common upscaler job is a scanned family photo. Expectations: a 600 × 800 scan enlarges cleanly to a 5 × 7 print (roughly 2×) and holds up as a framed 8 × 10 at the wall-viewing distance; the model will smooth heavy film grain (arguably fine), improve soft faces meaningfully, and can occasionally make one grandparent look subtly like a distant relative — facial detail is exactly where prediction is most seductive and most inventive. Compare the enlarged faces against the original side by side before ordering prints; a 200% zoom on the nose of the 1940 portrait is not a forensic question, it is a taste question, but the big features are where memory notices cheating.

A physical-scan note: the biggest quality gain on old photos is usually a better scan, not a bigger multiplier. A flatbed at 1200 dpi or a careful high-resolution phone copy of the print hands the upscaler real information about the emulsion instead of scanner-noise mush.