Why PicRemedy keeps image work in the browser

Who publishes PicRemedy

PicRemedy publishes this site. It is responsible for every page, for the model files the site serves and for corrections. The same name appears as the publisher in the page markup.

The site has no accounts, no comments and no uploads. Nothing on it is written by visitors, so every sentence is published by PicRemedy.

The one way to contact the publisher is the support address on the contact page. It is meant for browser and error reports, corrections and privacy questions. It is not meant for image files.

Why the site exists

Many people need to clean up or enlarge a picture and do not want to hand it to someone else's server. PicRemedy takes the other route. The model files travel to the visitor, and the image does not travel anywhere.

That choice has costs, and the site says so. A model has to download before it can run. A weak device may not run it at all. A browser tab has less memory than a server. Those limits are stated on the pages where they matter.

How the models were chosen and checked

Each model must clear a recorded check before its control appears. The checks are written down in files kept with the site's source, several of them dated receipts.

AI Clean

AI Clean is a compact network of 232,963 parameters. It was measured on a frozen internal experiment of 100 image patches, each 256 pixels across. The patches were damaged by JPEG compression, by WebP compression or by added noise.

Averaged over the three damaged groups, the output sat 0.91 dB closer to the undamaged original than the damaged input did. JPEG patches gained 0.71 dB, WebP patches 0.66 dB and noisy patches 1.35 dB. Patches that were already clean came back at 45.7 dB, above the 45 dB floor the model had to clear.

The receipt states its own limit. The experiment gates the packaged model. It is not a promise that your image will improve, and it does not show that original detail came back.

Neural Upscale

Neural Upscale is a larger network, about 34 MB, that produces a 2x or a native 4x output. The browser offers it only after it reproduces a recorded reference output within a fixed tolerance on the current graphics device. It accepts inputs of at most 512 by 512 pixels.

Background remover

The background remover uses Google's MediaPipe Selfie Segmentation weights under the Apache License 2.0. PicRemedy converted the file to another format and changed no weight value. On six reference cases the converted model matched the original interpreter to within 0.000014 in the worst case.

Before the control appears, the browser checks the file's length and its SHA-256 digest. It then reproduces a recorded mask fixture. A device that fails any step is told so.

Models the site refused

Widely recommended background-removal models were refused on licence grounds rather than on quality. The licenses page names them and gives the reasons. A permissive tag on a model file does not settle the terms of the data it was trained on, so the site checks both.

Editorial policy

The rules the site holds itself to are on the method page, together with a change log of dated entries.