AI photo enhancer: what it can fix, what it makes up, and when not to use it

What “enhance” actually does

A photo enhancer does two separate jobs, and it helps to keep them apart, because one is arithmetic and the other is guesswork.

The first job is light and colour. A dull, grey photo from a cloudy day, a yellow cast from indoor bulbs, a face lost in shadow against a bright window: these are fixed by stretching the range of brightness, shifting the colour balance and lifting the shadows. No model is needed. The information is in the file; it is just badly distributed, and the correction can be undone or adjusted without loss. Our tool does this part with sliders and an “auto fix” that sets them for you.

The second job is detail: making a small or soft photo larger and sharper. Here there is a hard limit that no software can get around. A photo 600 pixels wide holds a certain amount of information, and no process can recover detail that was never recorded. What an AI upscaler does instead is plausible invention: trained on millions of pairs of sharp photos and their blurred, shrunken copies, it has learned what edges, skin, hair, fabric and foliage usually look like, and it draws that in where the original only hints at it.

Done well, the result looks right, and for most purposes it is exactly what you want. But it is a drawing informed by the photo, not a recovery of the photo, and that distinction decides when to use it.

The model, and what it is good at

The enhancer on this site uses Real-ESRGAN, an open model built for restoring real-world photographs rather than clean computer graphics. It downloads once, about 19 MB, and then runs on your own device, part by part: the photo is cut into tiles, each tile is enlarged four times by the model, and the tiles are stitched back together at the size you chose (the same size, twice, or four times as wide and tall). A “strength” slider blends the model’s output with the original, so you can keep more of the real texture when the AI version looks too smooth.

Where it shines:

• Small photos from old phones, messaging apps and the web. A 640-pixel photo forwarded three times on WhatsApp comes back at 2,560 pixels with clean edges and no blocky compression artefacts. This is the most common use and the most satisfying.
• Slightly soft focus and camera shake. Edges are tightened and JPEG noise is removed.
• Old prints scanned at low resolution. Faces and text become legible; grain is smoothed.
• Screenshots and scanned documents destined for a presentation, where crisp edges matter more than fidelity.

Where it struggles:

• Heavy blur. A face that is a smear stays a smear, now a sharper-edged one. The model cannot know what was there.
• Very small faces. Below about 30 pixels across, a face has too little information, and the invented features may not look like the person. Enlarge a group photo and inspect each face before you trust it.
• Text in the photo. Letters are redrawn as plausible letter-shapes; a blurry “3” can become a sharp “8”. Never read numbers off an enhanced photo.
• Photos that are already large. The AI works on photos up to about 1,600 by 1,600 pixels (2.6 million pixels); bigger photos take minutes on a phone, so for those the tool fixes light and colour only, which is usually what a large photo needs anyway.

How many pixels a print needs

The usual reason to upscale is printing, and the arithmetic is simple once you know one number: a good print needs about 300 pixels per inch, which is 118 pixels per centimetre. Below about 150 pixels per inch, prints look soft at reading distance; large prints seen from across a room get away with less.

Print	Pixels needed at 300 ppi	Looks fine from	Minimum (150 ppi)
Passport photo, 35 × 45 mm	413 × 531	hand’s length	207 × 266
10 × 15 cm (4 × 6 in)	1,181 × 1,772	hand’s length	591 × 886
13 × 18 cm (5 × 7 in)	1,535 × 2,126	hand’s length	768 × 1,063
A4 (21 × 29.7 cm)	2,480 × 3,508	50 cm	1,240 × 1,754
A3 (29.7 × 42 cm)	3,508 × 4,961	1 m	1,754 × 2,480

So a 1,000-pixel photo prints well at 10 × 15 cm as it is, and enlarging it twice makes it good for A4. Enlarging four times takes it to 4,000 pixels, enough for A3, and that is where the model’s invention starts to show if you look closely: the print will be sharp, and the detail in it will be partly the model’s. The tool’s largest output is 4,096 pixels on the long side, which covers A3 at 300 ppi.

Check the result at 100 percent zoom before printing, not fitted to the screen, where everything looks sharp.

When not to enhance a photo

Passport, visa and ID photos. The United States, the United Kingdom, Canada, India and the Schengen states all require that the photo is not digitally altered, and an AI-enhanced face is altered by definition: skin is smoothed, edges are redrawn. Offices do reject photos that look retouched. Take a new photo in good light instead, and use the passport photo tools only to crop and size it; they change nothing inside the frame.

Evidence and documents. A photo of a car’s number plate, a receipt, a meter reading or a signature should not be enhanced before it is read or shared, for the reason above: the model redraws, and a redrawn digit is a different digit. Zoom into the original.

Photos you will be compared against. A profile photo enhanced to flawless skin and perfect sharpness sets an expectation; use the strength slider at a low setting, or fix only light and colour.

Photos that are already good. Enhancement adds nothing to a sharp 12-megapixel photo, and the AI step is skipped for photos that size anyway.

File size after enhancing

Enlarging a photo four times multiplies its pixels by sixteen, and a 300 KB photo can come out at 4 or 5 MB. If the result is for a form or a messaging app, run it through the compress tool afterwards and set the size you need; the sharpness survives compression far better than the blur did, because clean edges compress well and noise does not. If it is for printing, keep the large file: the print shop wants pixels, not kilobytes.

What stays private

The photo, the enlarged version and every intermediate tile stay in your browser. The only download is the model, and it comes from this site to you. For old family photos, scans of documents and anything you would not upload to a stranger’s server, that is the reason to use a tool that works this way, and the airplane-mode test described in the privacy guide shows it working with no connection at all.