About PhotoRescue

Most photos people want to fix are not old prints. They are recent photos that lost quality on the way: grainy night shots from a phone, pictures squeezed by a messaging app, product images too small for a store, thumbnails that need to become a print. PhotoRescue focuses on those problems.

There are three main models. Enhance removes noise, colour blotches and JPEG compression from phone photos and keeps the size. Upscale 4× makes an image four times wider and taller. Upscale 8× makes a small image eight times wider and taller, for posters and big screens.

Each model does one job, and you can combine them: when a photo is both damaged and small, run Enhance first and upscale the result.

Everything runs in the browser, so there is nothing to install and no graphics card needed. The processing happens on our GPU servers.

Some AI enhancers make photos look sharp by generating them anew: new skin, new hair, new letters. The result can look impressive and still be wrong, with a face that is not quite the person or a label that does not match the product. We think that is the wrong default for photos of real people and real things.

Our models first remove the damage and rebuild the shapes your pixels support. A second detail stage then adds fine texture only where noise, compression or enlargement wiped it out. A clean photo comes out of Enhance essentially unchanged, and text the models detect is protected rather than redrawn.

That texture is plausible, not a record of the scene. So we say plainly: the results are for viewing, sharing and printing, not for evidence, identification or reading text that was never legible. Our guide to faithful and generative upscaling explains the trade-off.

Our models learn from pairs of images: a clean photograph and a copy degraded the way real photos are degraded. Rather than generic blur and noise, we simulate what phones and sharing actually do. That includes sensor noise in low light, phone noise reduction, night-mode merging, app compression and resizing, calibrated on measurements from real devices.

We measure the models on held-out photographs that are kept separate from the training pairs, and inspect real phone captures as well. Photos you upload are never used to train AI, and they are deleted within 24 hours.

We would rather tell you up front than disappoint you after an upload. PhotoRescue does not repair scratches, tears or stains on old prints, does not colourise black-and-white photos, does not remove backgrounds and does not generate or replace faces. It cannot fix heavy motion blur or make unreadable text readable.

PhotoRescue is also not file-recovery software: it cannot bring back deleted photos from a phone, memory card or computer. It works on photos you still have.

Photos are personal, so we keep them for as short a time as possible. Uploaded files and results are deleted automatically within 24 hours. They are never sold and never used to train AI. You can use the free plan without an account or a card. The privacy policy lists every service involved and what each one stores.

The free plan gives you 3 credits a day, with no sign-up and no card, for JPG, PNG or WebP files up to 3 MB and 8 megapixels. One photo uses 1 credit, and Upscale 8× uses 2. Free results carry a watermark.

Pro costs $9.99 a month for 200 credits and Pro+ $19.99 a month for 500. Paid plans remove the watermark, accept uploads up to 20 MB and 30 megapixels and process batches of up to 50 images. See pricing for yearly prices.

Jorge Saez runs PhotoRescue from Rio de Janeiro, Brazil, and builds the models and service. The same operator is identified in the terms of service; there is no separate editorial team behind these guides.

The service is paid for by its subscribers, not by advertising, so we have no reason to sell or reuse your photos.

Contact: support@photo-rescue.com. The provider details are in section 1 of the terms of service.

Our method and limits explains paired validation, real phone checks and the saved GPU outputs used on this site. We compare the same crop on both sides and disclose when a source is deliberately downscaled or a result has not been tested as a physical print.

The tests found limits too: small text can come out wrong, bright edges can be affected by protection-mask differences, and some upscaled texture changes between runs. One attractive crop is not a guarantee for a whole photo. Our old-photo explanation separates cleaning a digital scan from repairing scratches, tears or faded colour.