Upscale an image to 2× or 4×

Super-resolution versus ordinary resizing

Ordinary resizing, the kind every image app does, spreads the existing pixels over a bigger grid and smooths between them. The result is larger but soft: edges blur and fine texture turns to mush. Super-resolution is different. A neural network that has seen many pairs of small and large images predicts what the sharper version probably looked like, and draws crisper edges and plausible texture.

The word that matters is plausible. The model does not recover detail the camera never captured. It makes a confident guess. On foliage, fabric, brick and landscapes the guess is usually convincing. On small faces and text it can look painted, especially at 4×.

So you can judge for yourself, the comparison view puts the model's result next to a high-quality bicubic resize to the same size. The label says which side is which.

Detail or Fast

Detail is the default engine. It is a 4.9 MB model that enlarges by 4×; for 2× output, the editor runs it at 4× and then scales down with a high-quality filter, which keeps the rebuilt edges. Choose it when the result matters.

Fast is a 240 KB model that works on brightness only, enlarging by 3×, and the editor then resizes to the scale you picked. It is lighter and quicker on slow devices, with a smaller gain over plain resizing. It is a reasonable choice for a quick check, or on an older phone where Detail would take a long time.

If you are unsure, run Detail first and look at the comparison. Switch to Fast only if the wait is too long on your device.

2× or 4×, and how big you can go

4× is offered only when the result stays inside what your device can hold: about 16.7 megapixels of output on iPhone and iPad, and 64 megapixels elsewhere. If your photo is too large for a scale, the editor tells you plainly instead of failing halfway. 2× is usually enough for a print or a sharper web image; keep 4× for small originals.

The image is processed in overlapping tiles, so the seams do not show. Progress reads “tile n of m”, and you can cancel at any point.

If you only need part of the photo larger, crop it first with Smart crop. A smaller input means fewer tiles, a shorter wait, and sometimes room for 4× where the full photo would not fit.

When the model is not used

If you decline the download, or the model fails on your device, you can still export an ordinary resize. That file is labelled “Standard resize”, never “AI upscaled”, so you always know which one you have.

What it will not do

• It sharpens and rebuilds plausible texture; it cannot recover detail that was never captured.
• Faces and text can look painted, especially at 4×.
• 4× is offered only when the output stays under about 16.7 MP on iPhone and iPad, or 64 MP elsewhere.
• A fallback without the model is labelled “Standard resize”, not an AI result.

Devices and downloads

Detail downloads 4.9 MB (BSD-3-Clause) and Fast downloads 240 KB (Apache-2.0); the editor shows the size first and starts after one press of the main button. Both run in a background worker, on WebGPU when your browser offers it and on WebAssembly otherwise. This is an advanced tool: tiled enlargement of a large photo is noticeably slower on WebAssembly than with WebGPU, and the tile counter shows how far it has got.

Advanced · faster with a GPU: A heavier network. It runs on any device, quickly where the browser offers WebGPU and slowly where it does not. See what is stored on this device.

Step by step

1. Open a small photo. Drop, paste or choose the image you want larger.
2. Pick scale and engine. Choose 2× or 4×, and Detail or Fast.
3. Run and wait for the tiles. Watch “tile n of m” and cancel if you change your mind.
4. Compare, then export. Check the model result against the ordinary resize before saving.

Questions about image upscaler