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.