How the upscaler works

The model is Real-ESRGAN General x4v3, a compact convolutional network of about 1.2 million parameters. It was trained on photos that had been deliberately blurred, resized, noised and JPEG-compressed, and learned to reverse that damage while enlarging by four times.

The exported model takes a fixed 128 × 128 pixel input, so your image is cut into overlapping tiles and each one is run on its own. Every tile is given a margin of real surrounding image, and only its centre is kept — so each output pixel is computed with genuine context around it and there is no visible grid where tiles meet. For 2×, the 4× result is averaged down rather than using a weaker model.

What it can and cannot do

An upscaler invents detail; it does not recover it. The network draws the sharp edges and fine texture a photo of that kind would most likely have had. On landscapes, products, renders and most photos that looks convincingly real. It cannot bring back information that was never in the file — tiny text will not become readable, and a blurred face will not become that person’s face.

Speed

With a GPU (WebGPU) — current Chrome and Edge on most computers — a typical photo takes seconds.
Without one, the same model runs on the CPU. It gives the identical result but is much slower: expect minutes for a 1 megapixel image. The progress bar shows an estimate once the first tiles are done.
The first run downloads the engine and the model (about 4.7 MB for the model). Your browser caches both, so later runs start straight away.

Privacy

The image is decoded and upscaled inside your browser and never uploaded. The page makes two downloads the first time you use it — the model file from this site, and ONNX Runtime, the open-source engine that runs it, from the jsDelivr CDN. Both fetch code; neither is sent anything about your image. See the privacy page for the full picture.

Limits

Images up to about 1 megapixel are accepted. The ceiling is the output, not the input: a 4× result has sixteen times the pixels, and past roughly 16 megapixels browsers — iOS Safari in particular — refuse to hold the canvas.

Credits

Real-ESRGAN by Xintao Wang, Liangbin Xie, Chao Dong and Ying Shan, released under the BSD 3-Clause licence; ONNX export by Qualcomm AI Hub. ONNX Runtime by Microsoft, MIT licence. The licence text ships alongside the model file.

Upscale an image Compress an image Remove a Gemini watermark