Skip to content

Latest commit

 

History

History
42 lines (26 loc) · 3.28 KB

README.md

File metadata and controls

42 lines (26 loc) · 3.28 KB
Raw

Neural Style Transfer

This repository contains PyTorch implementation of the original neural style transfer paper

How does it work?

The algorithm produces an image which combines style of the style image with contents of the content image using convolutional neural networks (the paper as well as this repo use VGG-19).

Examples

In all examples below a photograph depicting the Neckarfront in Tübingen, Germany is used as the content image.

a photograph depicting the Neckarfront in Tübingen, Germany

Below you can see a reconstruction of examples provided in the original paper.

The Shipwreck of the Minotaur by J.M.W. Turner an image combining style of The Shipwreck of the Minotaur and the Neckarfront in Tübingen The Starry Night by Vincent van Gogh an image combining style of The Starry Night and the Neckarfront in Tübingen Der Schrei by Edvard Munch an image combining style of Der Schrei and the Neckarfront in Tübingen Femme nue assise by Pablo Picasso an image combining style of Femme nue assise and the Neckarfront in Tübingen Composition VII by Wassily Kandinsky an image combining style of Composition VII and the Neckarfront in Tübingen

Running the scripts

If you are using conda, you can use environment.yml to recreate the environment that was used to develop code present in this repository.

Example usage of neural_style_transfer.py: python3 neural_style_transfer.py --style-img-path images/style/The_Starry_Night_van_Gogh.jpg --content-img-path images/Tuebingen_Neckarfront.jpg --height 240 --width 320 --save-freq 100 --beta 6000000 --alpha 0.1 --lr 0.1 --resize-style-img.

It is possible to try different parameters without the need to run neural_style_transfer.py multiple times. Simply run grid_search.sh $PATH_TO_STYLE_IMG $PATH_TO_CONTENT_IMG where $PATH_TO_STYLE_IMG is a path to the style image and $PATH_TO_CONTENT_IMG is a path to the content image. Note that in order to avoid clutter you have to manually edit grid_search.sh in a highlighted place to adjust the parameters that you want to try.