About this queue

PhotosUpscaler does one thing at two sizes. Given a directory, it applies one upward rule and one output format to every image in it and hands back an archive — no preview, no per-image control, no route named after a magnification, because a queue of five hundred files is not five hundred single-image sessions stacked up. Given one photograph, it opens a review studio instead, because a single photo is exactly where per-image judgement belongs.

The studio is built around one question: did the enhancement help, or did it make things up? It enlarges on this device — Lanczos3, ESPCN or Real-ESRGAN, in tiles — with JPEG and noise clean-up before and detail recovery after, and then judges the result against the same photo enlarged plainly: a split view, a loupe at 100%, and a risk map of where the network grew texture the source never had.

Super resolution in a browser tab is arithmetic that does not scale. One pass of the smaller network is a fifth of a second and one photograph can be a thousand passes, so the difference between a job that finishes over a coffee and one that does not finish at all is a factor a person cannot judge by looking at a folder.

So this site measures rather than promises. It runs eight real tiles from three of your own files on your own processor, takes the median, and quotes a range before it will start. Where the answer is bad it says the answer is bad, publishes the threshold it refused at, and puts the refused files through an arithmetic enlarger that finishes them in under half a second each. A queue here always reaches the end; sometimes it reaches the end having told you the model was the wrong instrument.

• There is no GPU in this path, and model passes use up to four CPU threads where the browser isolates the page (one where it cannot, as in Safari). A desktop application with a graphics card behind it will beat this by two orders of magnitude.
• It cannot make a picture smaller. The factor control has no value below one, and a resize to an exact pixel target is a different tool entirely.
• It cannot recover detail that was never recorded. It produces a plausible larger version, which is useful and is a smaller claim than the one this category usually makes.
• The queue cannot tell you whether each result is good enough for your listing; its report names soft sources, but judging a picture takes the review studio, one photograph at a time.
• The studio's risk map says where a network added texture, not whether that texture is true. A face, a label or a document still needs a person to look.

PhotosUpscaler is an independent product with one job: making photographs bigger and cleaner on the device they are already on, and being exact about what that took. It is built and run by a small team, it has no account system and no paid tier, and it shows no ads. Questions, corrections and a number that turned out to be wrong all go through the contact page, which sends only the text you type.

• Where the minutes come fromTile geometry read off both graphs, the per-pass figures, and the calibration that replaces them.
• Runs, warns, refusedThree bands, two thresholds, the halved iOS envelope and the arithmetic floor beneath all of it.
• PrivacyWhat leaves this device, what does not, and what is written to disk.
• TermsThe terms this is offered under, including the ones about accuracy.