AI image upscaling: what it can and can't actually fix in a low-res photo

A small, blurry photo — an old scan, a screenshot, a thumbnail that's the only copy left of something — gets stretched to a larger size and the result is exactly as blurry, just bigger. Simple resizing can't add detail that was never captured in the first place; a different category of tool tries to add it anyway.

Stretching vs. inventing detail

A basic resize (nearest-neighbor, bilinear, bicubic — the standard interpolation methods) makes an educated guess about the color of each new pixel based on its neighbors, which produces a smoothly blurred larger image but never anything sharper than the source. AI upscaling takes a fundamentally different approach: a model trained on huge numbers of high-resolution and deliberately-degraded low-resolution image pairs learns what plausible fine detail tends to look like, and generates new pixels that are a genuine best guess at detail rather than an average of existing ones.

Where the results genuinely hold up

Faces, text, and regular textures (brick, fabric weave, foliage) are the cases AI upscaling handles best, since these are exactly the patterns a model trained on millions of photos has seen the most examples of and can confidently reconstruct.

Where the invention starts to show

• Genuinely unique, irregular detail — a specific pattern that doesn't resemble anything common — sometimes gets "corrected" toward something more generic-looking, because the model is working from patterns it has seen before.
• Text in an unfamiliar font or a language the model wasn't trained on much can come out looking like plausible-but-wrong characters rather than the actual ones.
• Heavily compressed or very low-resolution sources give the model less real signal to work from, so more of the output is the model's invention rather than a recovery of lost detail.

For anything where accuracy matters more than a pleasing result — a document you need to actually read, evidence, a historical record — treat an upscaled image as an aid to viewing, not a substitute for the original source.