This is a tensorflow implemenation of Deep Image Prior.

Setup

Install python 3.6, tensorflow (tested on 1.5.0+), numpy.
Run it: python deepimg.py
Every 100 iterations, the current image is written to the output directory.
- The input image will be blurred and written to output/corrupted.png. This is the starting image that the model attempts to sharpen.

Known issues and discrepancies

This only implements super resolution.
This uses a Gaussian blur rather than a Lanczos2 kernel for the downsampling operator in E.