An implementation of https://dmitryulyanov.github.io/deep_image_prior for tensorflow.
Switch branches/tags
Nothing to show
Clone or download
Latest commit e707a80 Mar 23, 2018

README.md

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.

Example

Blurred pupper 2000 iterations pupper 4000 iterations pupper