Sandbox · PHP + a neural network in your browser

Image upscaler: 4x enlargement, three ways

We take a small image, make it four times larger and put the results side by side: classic bicubic interpolation, a direction-aware PHP pipeline running on the server (DCCI), the style-preserving MMPX for pixel art, and a neural network (Real-ESRGAN) that runs entirely in your browser. And we measure them, not just look at them.

What upscaling can and cannot do

The neural output is a suggestion, not evidence. Whatever the small image no longer contains, the network invents: it paints plausible texture, edges, even letters. That is why it cannot "restore" a blurry licence plate, face or caption.

Sample image

NASA photo: the Artemis I rocket inside the Vehicle Assembly Building (NASA/Frank Michaux, public domain).

Drop an image here, or choose one. Your browser decodes the image into raw pixels and shrinks it to at most 256x256. The server only receives those pixels, stores nothing and never runs an image decoder on them. The neural run stays entirely on your device.

Methods (up to 4)

Input (pixelated)

Bicubic

DCCI (PHP)

MMPX (PHP)

Neural: Real-ESRGAN

Neural: anime model

Original large image

The first run downloads about 8,6 MB (the network and the runtime); after that your browser keeps it. It only starts when you press this button.

View

The link contains the sample, the methods and the view (never your own image).

Left side

Right side

BicubicDCCI (PHP)

Bicubic

DCCI (PHP)

Timing

The server times itself step by step (Server-Timing header); the browser measures the full round trip. The neural run is timed on your own device.

No measurement yet.

Models: Real-ESRGAN (Xintao Wang, BSD-3-Clause), runtime: ONNX Runtime Web (MIT). Algorithms: DCCI (Zhou, Shen, Dong, 2012), MMPX (McGuire, Gagiu, 2021).