Enhance old photos and clean up grainy scans

Why old photos look grainy and soft once digitized

Part of the roughness is in the photo itself. A developed film image is made of tiny grains of silver or clouds of dye, and faster film has coarser grain, so enlarging a small negative into a print enlarges that texture too. Many family cameras also had simple fixed-focus lenses that were never very sharp, especially toward the corners.

More comes from the copy. Old scans were often made at low settings to keep files small, then saved as heavily compressed JPEGs that add blocky patches and smear fine texture. A phone photo of a print brings its own problems: glare on glossy paper, uneven light, a skewed shape when the phone is tilted, and the phone's own noise reduction and sharpening.

Time adds a third layer. Color dyes fade at different rates, so many prints drift toward red, magenta or yellow, and the paper collects scratches, creases and stains. To tell which you have, zoom to full size and study plain areas such as sky or a wall:

• Fine speckle everywhere, even in clear sky: film grain or scanner noise
• A regular pebbly or honeycomb pattern: textured photo paper
• Square blocks and fuzzy halos around edges: heavy JPEG compression
• A washed-out patch or one darker side: glare or uneven light in a phone capture
• Rows of tiny dots: a picture printed in a newspaper, magazine or yearbook

What an AI enhancer does to an old photo

Every upload goes to Google's image AI under one brief: upscale it, clean up noise and JPEG blocks, sharpen fine detail and edges, and change nothing about the subject, composition, framing or colors. On a scan of an old print, that usually means film grain and scanner noise settle down, the outlines of faces, hair and clothing are redrawn more crisply, and compression blocks smooth out.

The model does not uncover detail hidden under the grain. It draws new detail that fits what it sees, based on patterns learned from other images, so the fewer pixels a face has, the more of it is a guess. Check faces at full size before you share the result.

Grain is also part of how film looks. There is no strength setting, so an enhanced scan can come out cleaner and smoother than you expect. If the original feels more like the photo you remember, keep it.

Enhancing an old photo is not restoring it

Restoring an old photo usually means repairing it: filling tears, removing scratches and stains, hiding crease lines, rebuilding missing corners and correcting faded color. Pixelperfect does none of that. The engine is told not to add or remove content and to keep the colors it receives, so expect damage to stay where it is, along with any faded cast or sepia tone.

Colorizing is a separate job again. A black-and-white photo only recorded brightness, so any colorized version is a guess by software or an artist, not a recovery of the real colors.

Fix damage with the healing or clone tools in a photo editor, or hire a restoration specialist for photos that matter most. A color cast often responds to an editor's white balance or auto color control.

Get a better scan, then crop to what matters

The cleaner your starting file, the less the model has to invent. A flatbed scanner makes that easy (flat print, even light, no glare), and a phone works too if you shoot straight-on in soft, even light. Either way, crop off the border and background so every pixel belongs to the photo.

Size matters as well. Your browser shrinks any image over about 1.2 megapixels, or over 1600 px on its longest side, before upload. The enhanced version is about four megapixels (a square photo comes out near 2048 x 2048 pixels), or about one megapixel when the image you upload is under about a third of a megapixel. Small and medium scans come back larger; a big, high-resolution scan comes back smaller than it went in.

That is why cropping helps. In a whole group photo, each face may end up a few dozen pixels wide. Crop one face from a high-resolution scan instead, keeping the crop under about 1000 x 1000 pixels, and that face usually reaches the model at full scan resolution, not shrunk with the rest. A crop cannot add detail the scan never captured, so if a face is tiny even in the scan, rescan at a higher setting first.