How to upscale a low-resolution image the right way.

Blow a photo up past its real size and it goes soft or blocky — the pixels needed to make it sharp were never there. Modern AI upscalers do something cleverer: they invent plausible detail to fill the gap. Here's what that can fix, what it can't, and how to do it without your photo ever leaving your device.

Last updated 2026-07-12

Why 'just make it bigger' looks bad

Every image is a fixed grid of pixels. When you enlarge a 400×300 photo to 1600×1200, the software has to conjure fifteen out of every sixteen pixels in the new grid — and the classic methods do it by guessing from the neighbours. Nearest-neighbour just copies the closest existing pixel, which gives hard, blocky stair-steps. Bilinear and bicubic average the surrounding pixels instead — bicubic weighs a 4×4 block of them — so edges turn smooth rather than jagged. But 'smooth' here is a polite word for blurry. Crucially, none of these methods add information; they only spread the original pixels across a bigger canvas. That's why a stretched thumbnail never becomes a real high-resolution photo. It just becomes a larger thumbnail.

What an AI upscaler actually does

A modern AI upscaler — the technique is called super-resolution — works from a completely different angle. Models such as Real-ESRGAN are neural networks trained on vast numbers of image pairs: a sharp original next to a deliberately wrecked copy carrying blur, noise and JPEG compression. Over that training the network learns what real edges, textures and faces are supposed to look like. Feed it a soft, low-res input and it predicts a high-res version — crisp eyelashes, clean lettering, defined brickwork — and cleans up compression noise on the way. The word doing the heavy lifting is predict. The model isn't uncovering pixels hidden in your file; it's generating brand-new ones that are statistically plausible for what it sees. Done well, the output genuinely looks clearer, not merely bigger.

The limit no upscaler can beat

This is exactly where the crime-drama 'zoom and enhance' fantasy breaks down. An upscaler can only add detail that fits its training; it cannot retrieve information your camera never recorded. If a face is six pixels across, the model will confidently draw a detailed face — but it's an educated guess, and it may not resemble the actual person. The same trap waits with a blurred licence plate or a serial number: you'll get crisp-looking characters with no guarantee they match what was really there. So AI upscaling is wonderful for making a picture look good, and a bad idea anywhere accuracy carries weight — evidence, medical images, or reading text you couldn't almost read already. Treat the result as a convincing reconstruction, never as recovered truth.

Where it works well (and where it doesn't)

Upscaling earns its keep on old family photos, low-resolution scans, product shots, small logos and web images that were saved too small — cases where you want a cleaner, larger copy and nothing factual rides on the exact pixels. Portraits and natural textures like skin, hair, foliage and stone tend to come out best, simply because that is what the models have seen most. It struggles elsewhere: a photo-trained model can smear stylised or anime line-art, it can't rescue text below a legible size, and it has little to do with images that are already sharp. Scale matters too. 2× is a safe, natural bump; 4× is impressive but far more likely to invent detail you'll catch on close inspection. Because the whole thing runs locally and for free, the pragmatic move is to try both factors on your own image and keep whichever holds up at full zoom, rather than guessing in advance which one will win.

Upscaling up vs. resizing down

It helps to be clear about which problem you actually have. Shrinking an image — resizing down — is the easy direction: you're discarding pixels you don't need, and an ordinary resize stays crisp because it never has to guess. Enlarging is the hard direction, with nothing to discard and everything to invent. So if your real goal is 'fit this into a 600-pixel slot' and the original is already bigger, don't upscale at all — a plain resize does the job without hallucinating anything. Save AI upscaling for when the source honestly lacks the resolution you need. Our resize tool at /resize-image/ covers the shrink-and-fit case; the upscaler at /upscale-image/ is for when there simply aren't enough pixels to start with.

It all runs in your browser

Upscaling is the kind of job you'd rather not hand to someone else's server — it's usually old family photos, ID scans, documents and other personal images. GotConvert's upscaler doesn't upload any of it. The first time you use it, a compact Real-ESRGAN model — the roughly 5 MB realesr-general-x4v3, tuned for real-world photo degradation — downloads into your browser, and from then on every pixel is processed locally with WebAssembly. Your image never leaves the tab: no queue, no watermark, no account. The honest trade-off is that speed depends on your own hardware, and very large 4× jobs are capped so a phone doesn't run out of memory — a fair price for the file staying entirely yours.

Frequently asked

Can AI upscaling recover the real detail in a blurry photo, like reading a licence plate?

No. An AI upscaler synthesises plausible detail from what it learned during training; it can't recover information the camera never captured. It will happily produce sharp-looking digits or a detailed face, but there's no guarantee they match reality — so never rely on upscaled output as forensic or legal evidence.

What's the difference between upscaling and just resizing?

Resizing down discards pixels you don't need and stays crisp. Upscaling has to invent pixels that were never there. A basic resize (bilinear or bicubic) can enlarge too, but it only spreads existing data, so it looks blurry; an AI upscaler predicts new detail instead. If you only need the image smaller, resize — don't upscale.

Should I choose 2× or 4×?

2× is the safer choice: the jump in size is smaller, so any invented detail stays subtle and is less likely to look wrong. 4× is dramatic but reconstructs far more, so artifacts and made-up detail are easier to spot on close inspection. If you need to go bigger than a single 4× pass, upscale once and run the result through again rather than expecting one giant leap to look clean.

Will upscaling make tiny, unreadable text legible?

Usually not. If the text is already close to readable, upscaling can sharpen it just enough to help. If it's only a smudge of a few pixels, the model guesses at letter shapes and can produce confident but wrong characters. Don't trust reconstructed small text — the sharpness is real, the accuracy isn't guaranteed.

Does my image get uploaded to a server?

No. The upscaler downloads a small Real-ESRGAN model into your browser and runs entirely on your device with WebAssembly. Your photo never leaves the tab — there's no upload, no account and no watermark — which matters when you're enlarging personal photos, ID scans or documents.