SRCNN online

Unofficial online demo for "Image Super-Resolution Using Deep Convolutional Networks" (often known as SRCNN).

Super-resolution problem tries to upscale the image so that perceived loss of quality is minimal. For example, after scaling with bicubic interpolation it is apparent that some pixels are just smudged together. The question is: can an AI do a better job?

While this technique is old (published in 2015), it makes for a good PyTorch/ONNX demo.

The model was trained in PyTorch for 2 minutes using some random images from the web. It was then exported into ONNX file. Unfortunately, ONNX runtime on the web has errors that prevent using GPU backends (WebGPU, WebGL). On the CPU, it should take ~15s to process the test image.