Upscale a Low-Resolution Photo — Free, Fast & Private
"Low resolution" is a relationship, not a property, and naming both halves of it makes the decision easy. A 600-pixel photograph is high resolution for an avatar, adequate for a small blog illustration, marginal for a feed post and hopeless for an A4 print. So before reaching for any tool, write down two numbers: the pixel dimensions the file actually has, and the pixel dimensions the destination actually wants. The gap between them is the only thing that decides whether you need a 2x rebuild, a 4x rebuild, or nothing at all — and quite often the honest answer is the third one, because the target is smaller than people assume and the file was fine all along. When it is not, the on-device upscaler is the tool that closes the gap.
What each starting size can realistically become. Under about 300 pixels on the long edge, a 4x run produces 1200 pixels, but the model is being asked to invent fifteen pixels for every one it can see; expect something that looks drawn rather than photographed, usable as a thumbnail and unconvincing at any size where a viewer can inspect it. From 300 to 600 pixels you land between 1200 and 2400 — a genuinely usable web image, and the range where the difference between reconstruction and a plain stretch is most dramatic. From 600 to 1000 pixels is the sweet spot: enough real structure for the model to build on, and an output between 2400 and 4000 pixels that covers most screen and small-print targets. Above roughly 1500 pixels, check the target before running anything, because you may already be there. These are practical bands rather than hard switches, and the picture matters too — a landscape survives aggressive enlargement far better than a group portrait, because nobody scrutinises a hillside the way they scrutinise a face.
Where low-resolution copies come from, and why the fix is often not a tool. Most files that arrive too small were not shot that way. Messaging apps resize attachments on send, email clients offer a "smaller size" option that is frequently the default, social platforms serve a downscaled render rather than the upload, and screenshots capture whatever the screen was showing rather than the file behind it. Each of those produces a reduced copy while the original sits untouched somewhere else. Five minutes checking a camera roll, a cloud backup, the original email thread or the sender device is worth more than any reconstruction, because a larger original holds detail that was genuinely recorded. Only once that search has failed does an upscaler become the right answer rather than the convenient one.
How the rebuild works, and what it will not do. The free path runs realesr-general-x4v3, a compact Real-ESRGAN network: 4.9 MB, verified against a pinned SHA-256 hash, executed in your own browser through WebGPU or WebAssembly, tile by tile so memory stays bounded whatever the image size. It was trained on pairs of sharp images and shrunken degraded copies, so instead of averaging neighbouring pixels it predicts the texture that would plausibly have produced each small patch. That is a real improvement in appearance and it is not a recovery of information: a face that dissolved into a handful of pixels comes back smooth and plausible, not identifiable, and a distant sign comes back sharp and wrong. Nothing from an upscaler should be used to identify a person, read a plate or support a claim. The AI image upscaler hub sets out the model and the tiling in full, and enlarge an image without losing quality takes the same ground from the quality-loss angle.
The limits you will meet, stated as numbers. Free accepts inputs up to 1080px on the long edge and delivers up to 4320px at 4x, with a small "minipx.com" watermark in one corner; it includes 2 AI runs shared across the three on-device AI tools, after which a cooldown starts at 60 minutes and lengthens by 30 minutes each further time it is spent, capping at 24 hours and clearing after 7 days without a run. Pro removes the watermark, makes on-device runs unlimited, and raises the input caps to 2048px at 4x and 4096px at 2x, both ceilinged at 8192px of output. Anything above the cap is downscaled to fit before the run and the tool says so rather than silently changing your dimensions. Browsers impose their own canvas memory ceilings on top of all this — iOS Safari is the strictest — so a phone can cap the output below the plan limit, and again the tool reports it.
Cloud HD for the sources the small model cannot carry. Pro accounts see a switch beside the run button, off until you turn it on. Switched on for a run, that one photo goes to our GPU where a JPEG pre-clean pass runs before HAT-L — a far larger super-resolution model — rebuilds it at 4x. The intake is 25 MB and 4 megapixels, output reaches up to 16× the pixels of the input — a 4 MP photo comes back at about 64 MP, and a full 4 megapixel source takes four to five minutes, which is stated before you commit. Each run spends one of the 10 HD runs in the month, inside a wider allowance of 150 runs on Pro Monthly or 200 on Pro Annual, resetting on the first with no rollover and a ceiling of 30 runs an hour; a failed run is not charged. The case for spending one is a small, compressed, irreplaceable source. The case against is a clean photo that the on-device model already handles, which is most of them.
Order of work, and where the file goes. Crop before you enlarge with the cropper, so no part of the run is spent on an area you will trim. If the source is visibly blocky rather than merely small, start at fix a pixelated image, which separates the three causes of blockiness and says which are treatable. Compress last, in the image compressor, since output is a lossless PNG that will be far heavier than a web page wants. On privacy: the free path processes everything on your own hardware, and once the model is cached the tool works offline — the network panel will show the model coming down and nothing image-shaped going up. The single exception is a Cloud HD run you switch on yourself, where that one photo is uploaded encrypted, processed in the United States on Cloudflare R2 and a Modal GPU, deleted when the run finishes, and in all cases within 24 hours, and is never used to train anything.