AI Image Upscaler — Free, Fast & Private
MiniPx's AI Image Upscaler enlarges images 2× or 4× using realesr-general-x4v3, a compact super-resolution model from the Real-ESRGAN project. The whole thing runs in your browser: the 4.9 MB model downloads to your device once, is verified against a pinned checksum, and every pixel of inference happens locally through WebGPU (or WebAssembly on devices without it). Your photo is never uploaded — not to us, not to anyone. After the first run the model is cached, and the tool works offline.
How AI upscaling actually works. Traditional enlargement — bicubic or bilinear interpolation, what happens when you just stretch an image — averages neighbouring pixels, which is why stretched images look soft. A super-resolution network takes a different route: it was trained on millions of pairs of sharp images and their artificially degraded, shrunken counterparts, and it learned to predict what plausible high-resolution texture looks like for a given low-resolution patch. Edges come back crisp, surfaces get believable grain, hair and foliage regain structure. The architecture here is SRVGGNetCompact, a lean stack of convolutions designed by the Real-ESRGAN authors specifically to be fast enough for real-time use — which is what makes running it in a browser practical at all.
What it cannot do matters just as much, and most upscaler marketing skips it. The model synthesises plausible texture; it does not recover information that was never captured. A licence plate that is an unreadable smudge at 200px stays unreadable at 800px — just a sharper smudge. A face that dissolved into blur will come back smoother, not identifiable. A genuinely out-of-focus photo stays soft, because softness is what is actually in the pixels; the model sharpens sampling artifacts, not optics. This is texture synthesis guided by learned priors, and everything it "adds" is a statistically plausible invention, not a recovery. For photos of documents, faces you need to recognise, or anything evidentiary, treat the output as an enhancement of appearance, never of information.
Under the hood: tiled inference. Super-resolution is memory-hungry in a way most browser AI is not — the output tensor is sixteen times the input's pixel count at 4×, so running a full image through the network in one pass would crash a tab on anything past thumbnail size. MiniPx processes the image in 512px tiles (256px on WebAssembly) with a 16px overlap pad, the same tiling strategy the Real-ESRGAN reference implementation uses. Each tile runs through the network alone and is stitched into the output canvas with the padding cropped away, so seams land pixel-exact and peak memory stays flat no matter how large the image is. That is what the tile counter in the progress bar is showing you.
2× or 4×? Both settings run the same 4× network — the 2× option downsamples the result on the way out, which softens the synthesised texture slightly and halves the output dimensions. Pick 4× when the source is genuinely small: thumbnails, old photos from early phone cameras, screenshots, game assets under about 1000px. Pick 2× when the source is already reasonable and you want a safer, subtler enlargement — over-upscaling a decent image can push the texture synthesis into a slightly artificial, over-etched look. On the Pro tier, 2× also accepts larger inputs (4096px vs 2048px), since the output ceiling is what actually constrains memory.
Photos vs graphics. The general-purpose model is at its best on photographs: skin, fabric, foliage, product shots, architecture. It also does well on illustrations and flat-colour graphics, where clean edge reconstruction is most of the job. Text in images sharpens legibly if it was already borderline-readable, but the model will not reconstruct lost letterforms. For logos and icons that exist as vectors somewhere, re-exporting from the vector source will always beat any upscaler — use this tool when the raster copy is all you have. Screenshots of UI upscale surprisingly well at 2×; pixel art is a matter of taste, since the model smooths deliberate pixelation rather than preserving it.
Free and Pro, stated plainly. The free tier accepts inputs up to 1080px on the long edge — enough for a 4320px output at 4×, which covers most listing, thumbnail, and social uses — and stamps a small "minipx.com" watermark in the corner. Pro removes the watermark and raises the input cap to 2048px for 4× and 4096px for 2×, both of which top out at 8192px output. If your image exceeds the cap, it is downscaled to fit before upscaling and the tool tells you so explicitly — no silent surprises in the output dimensions. Browsers impose their own canvas memory ceilings too (iOS Safari is the strictest), and on mobile that can cap output size below the plan limit; again, the tool says so when it happens.
Privacy is structural, not a promise. Most online upscalers receive your image on their servers, run inference on their GPUs, and send the result back — which means your photo exists, at least transiently, on hardware you do not control. MiniPx inverts that: the model comes to you. There is no server-side image path to secure because there is no server-side image path at all. You can verify this yourself — open your browser's network inspector while upscaling and watch: the only download is the model file, and nothing image-shaped is uploaded. It also means no quota system, no queue, and no account: the marginal cost of your upscale is borne by your own device.
When upscaling is the right tool — and when it isn't. The clearest case is a hard resolution floor you cannot meet: a marketplace that wants 1600px for zoom, a print shop asking for 300 DPI, a thumbnail spec your source falls short of. Print is straight arithmetic — a 6×4 inch print at 300 DPI needs 1800×1200px, so a 900×600 photo needs exactly a 2× upscale, and an A4 poster at 150 DPI needs about 1754px on the short edge. If your image already exceeds the target, don't upscale for "extra sharpness": you gain nothing and risk the over-processed look. And if the target is only modestly above the source — say 10-20% — plain resizing in the resizer is often indistinguishable and instant. AI upscaling earns its keep in the 2×-4× range, where interpolation visibly falls apart.
A practical workflow note: upscaling and compression pull in opposite directions, and order matters. Upscale first, at maximum quality, then run the result through the compressor to hit your target file size. Compressing first hands the upscaler JPEG artifacts to enlarge, and it cannot reliably tell artifact texture from real texture. Pairs naturally with the background remover for product shots — cut out first, upscale the cutout second — and with the resizer when a platform wants one exact dimension.