Enlarge Image Without Losing Quality — Free, Fast & Private
The phrase hides an assumption worth unpicking. Nobody enlarging a photo is worried about the pixels they already have — those survive any enlargement untouched. The worry is about the pixels that have to be invented to fill the space between them, and the way ordinary software invents them. Bicubic and bilinear resampling, which is what every image editor and every browser reaches for by default, computes each new pixel as a weighted average of the four or sixteen nearest originals. Averaging is blurring. Do it once across a whole picture and every hard edge in the frame acquires a soft halo, every fine texture flattens toward its local mean, and the file that comes out is unmistakably bigger and unmistakably worse. So "without losing quality" is really a request for a different filling strategy, not for a magic that keeps a small picture small.
What a super-resolution network does instead. The model on this page is realesr-general-x4v3 from the Real-ESRGAN project — a compact SRVGGNet, 4.9 MB, integrity-checked against a pinned SHA-256 hash before it is allowed to run. It was trained on enormous numbers of image pairs: a sharp original, and the same original shrunk and degraded. From that it learned a mapping in the opposite direction, so when it meets a soft, small patch it does not average anything; it predicts what full-size texture would most plausibly have produced that patch. Edges come back as edges. Hair regains strands, fabric regains weave, foliage regains individual leaves. Run at 4x, the output holds sixteen times the pixels of the input, and the added ones look like photography rather than like arithmetic.
The honest ceiling, stated before you spend time on it. Prediction is not recovery. If detail was never captured — a face that dissolved into six pixels, a sign too far away to read, a photo that was out of focus at the shutter — the model produces a confident, plausible, wrong answer, and it will look sharper while being no more informative than before. Treat every upscaled file as an improvement in appearance, never in evidence. The same limit sets the practical range: 2x and 4x are the two factors offered, and there is deliberately nothing above 4x. Beyond that the ratio of invented to observed pixels tips far enough that the picture reads as a rendering. If your target needs more, the answer is a bigger source, not a second pass over the first result — the upscaler hub sets out why the factor list stops where it does.
What the free path gives you, in numbers. On the free tier the input cap is 1080px on the long edge, which at 4x produces output up to 4320px — larger than a UHD frame — and the file carries a small "minipx.com" watermark in one corner. Free also 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 the allowance is spent, capped at 24 hours and cleared after 7 days without an AI run. Pro removes the watermark, makes on-device runs unlimited, and raises the caps to 2048px in at 4x and 4096px in at 2x, both ceilinged at 8192px of output. Anything larger than the cap is downscaled to fit first, and the tool says so rather than quietly changing your dimensions. Sibling reading: the AI image upscaler hub documents the tiling and the model in full, and upscale a low-resolution photo covers what to do when the source is genuinely tiny.
Cloud HD is the optional heavier answer. Pro accounts see a switch beside the run button, off until switched on. Turning it on for a run sends that one photo to our GPU, where HAT-L — a far larger transformer-based super-resolution model — runs at 4x with a JPEG pre-clean pass first. 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 the progress bar tells you before you commit. It spends one of the 10 HD runs in the month, inside the wider allowance of 150 runs on Pro Monthly or 200 on Pro Annual; a run that fails is not charged. On a clean, well-exposed source the free model is already close, and the switch is not worth spending an HD run on. On a small, compressed, awkward source the gap is obvious.
Sequence matters more than settings. Crop before you enlarge, using the on-device cropper — pixels spent reconstructing scenery you are about to trim are pixels wasted, and cropping afterwards throws away the most expensive part of the run. Enlarge second, at the highest factor the source can honestly support. Compress last, in the image compressor, once the picture is final: a 4x PNG is routinely 10 to 25 MB, which is right for an archive and wrong for a web page. Doing it in the other order hands the model compression blocks to enlarge, and it cannot reliably tell a block from a brick. If the source is blocky to begin with, start at fix a pixelated image instead.
Where your photo goes. The free path is entirely local: the model file downloads to your browser and every pixel of inference happens on your own hardware, which you can confirm by watching the network panel during a run — the model comes down, nothing image-shaped goes up. After the first run it is cached and the tool works offline. The one exception on this page is Cloud HD, and only if you are on Pro and turn it on yourself: that single photo is uploaded encrypted, processed in the United States on Cloudflare R2 storage and a Modal GPU, deleted when the run finishes, and in all cases within 24 hours, and is never used to train anything. Every other MiniPx tool — the resizer, the converters, the compressors — has no upload path at all.