Enhance phone photos

A phone sensor is tiny. In good light that hardly matters. Indoors and at night the camera raises its sensitivity (ISO) to get a usable exposure, and the sensor's random noise is amplified with the picture. You see it as grain, and in the shadows as blotches of red, green and purple that are not part of the scene.

The phone then tries to hide that noise. Its built-in noise reduction blurs fine texture together with the grain. That is why night photos often look smeared or painted: hair turns into a smooth mass, foliage into green watercolour, brick into flat orange. Night modes that merge several frames help, but they can add smearing and ghosting when anything moves.

The last step usually happens when the photo is shared. WhatsApp, Messenger and most social apps shrink the picture and save it again as a compressed JPEG. Compression throws away the fine detail that was left and adds square blocks and halos around edges. Each forward can add another round. Our page on fixing WhatsApp photo quality explains this step in detail, and why night photos are grainy covers the camera side.

Enhance runs two neural networks in a row. The first is a restoration model. It removes sensor noise, colour blotches, smearing from phone noise reduction and JPEG compression in one pass, and keeps real edges where they were. The second is a detail model. Where noise or compression wiped out fine texture, such as the weave of fabric or the needles of a pine tree, it adds natural texture back so the result does not look plastic.

The detail model only works where the photo needs it. If your photo is already clean, Enhance leaves it essentially as it is; it does not invent detail that is not needed. Text it detects in the photo, such as signs and labels, is left to the restoration step, which removes noise and compression but does not redraw letters.

Both networks were trained on pairs of a clean photograph and the same photograph passed through a simulation of real phones. The simulation was calibrated on measurements of real devices: noise from budget and older phones, night-mode merging, and the resize-and-compress rules of messaging apps. That is why Enhance recognises phone damage that a generic denoise filter treats as detail.

The output has exactly the same width and height as your upload. Enhance does not enlarge, crop or relight the photo, and it does not change white balance or exposure: a warm night scene stays warm. If you also need a bigger picture, run Upscale 4× on the enhanced result.

Enhance makes the biggest difference when the damage is noise or compression, not missing focus:

• Photos taken in low light, at night or indoors that came out grainy, smudgy or soft: city lights, restaurants, concerts, birthday parties.
• Pictures that lost quality after being sent through WhatsApp or another chat app, or saved many times, where the original is gone.
• Photos from older or budget phones, whose small sensors are noisy even indoors.
• Product and listing photos taken in a dim shop or garage, before they go online.

Enhance repairs the picture; it does not read it. It will not make unreadable text readable, recover a licence plate or sharpen a document. Very small or dense print can occasionally be missed by the text detection, so for documents where every character matters, check the result or keep the original.

The fine texture it adds is plausible, not a record of the scene. Do not use the results as evidence or for identification, for example of faces, plates or serial numbers. And an already sharp, well-exposed photo will barely change: Enhance is a repair tool, not a filter that makes every photo look different.

Enhance was built for photographs. On screenshots and graphics it also removes compression blocks, but flat areas can pick up faint texture, so compare the result with the original.

The model can only work with what is in the file, so a better input gives a better output:

• Use the least-forwarded copy you can find. The photo in the sender's camera roll beats one forwarded three times.
• Avoid screenshots of photos. A screenshot adds a resize and interface scaling on top of the original damage.
• Upload the photo before you edit it. Filters, heavy sharpening and brightness boosts make noise harder to separate from detail.
• Enhance first, then upscale. Enlarging a noisy photo enlarges the noise, and noise is harder to remove once it is spread over more pixels.
• Check faces and any text at 100 % zoom before you print or publish.

Many AI enhancers are fully generative: they redraw skin, eyes and texture from scratch. That can look impressive at a glance and wrong on a closer look, especially on a person you know. Enhance is tuned to repair what the pixels support. It does not beautify faces, change expressions or add objects. Where information is truly gone, it leaves the area clean and natural instead of inventing it. Our guide to faithful and generative upscaling explains the difference.

Every example on this page and in the gallery is a real output of the model. The damaged inputs were made with our phone simulator from clean photos, and each example says so.