Two kinds of upscaling
Classical upscaling interpolates: it estimates each new pixel from its neighbours. It is instant, predictable and never invents anything, but it cannot add detail, so a 4× enlargement looks smooth rather than sharp.
AI upscaling uses a neural network trained on pairs of small and large photographs. Real-ESRGAN, the model used here, redraws edges, fine texture and text strokes at four times the size. It runs on your device after a one-time 4.6 MB download and takes seconds to minutes depending on the picture and whether your browser offers WebGPU.
Getting the best enlargement
- Start from the largest, least-compressed original you have; every enlargement magnifies JPEG blocks too.
- Up to 2×, Lanczos-3 is usually indistinguishable from AI on clean photos and costs nothing.
- Above 2×, compare Mitchell with AI detail ×4 in split view. The AI wins on edges and texture; Mitchell wins when you need zero invented detail (evidence, scientific images).
- For a precise target (a print, a 4K screen), set the size directly — the AI pass is followed by an exact resample to your numbers.
Questions
Is 2× AI possible if the model is 4×?
Yes. The network enlarges 4×, then the result is reduced to your exact target with Lanczos-3 in linear light. Reducing a 4× result to 2× keeps the reconstructed detail and averages away most model artefacts.
How long does AI upscaling take?
The picture is processed in tiles and the progress bar counts them. With WebGPU a 1-megapixel photo usually takes a few seconds; on the WebAssembly fallback (no GPU) budget around two seconds per tile. The panel shows an estimate before you start.
Will upscaling fix a blurry photo?
Partly. AI detail ×4 sharpens soft edges and removes mild JPEG artefacts as it enlarges, but it cannot restore detail lost to heavy motion blur or focus errors.
What does “linear light” mean, and why is it on by default?
Pixel values in a JPEG or PNG are gamma-encoded: value 128 is not half as bright as 255, it is about a fifth. Averaging those numbers directly — what most resizers do — darkens fine bright detail and muddies colour edges. With linear light on, every pixel is converted to physical light before it is weighted and converted back afterwards. The test chart’s one-pixel checkerboard shows the difference: reduced in linear light it matches the light grey patch beside it; reduced the naive way it matches the dark one.
Is my picture uploaded anywhere?
No. Decoding, scaling and encoding all run in this browser tab, and the scaling engine runs in a Web Worker with no network code in it. The only download the site can make for you is an AI model file, and only after you press the button that names its size — that request contains no picture.