GPDM

Original Pytorch implementation of the GPDM algorithm introduced in

"Generating Natural Images with Direct Patch Distribution Matching"

Accepted to ECCV 2022

Live-demo | Paper

Video presentation

Run GPDM:

Reshuffling

$ python3 main.py data/images/SIGD16/7.jpg

Input	Output

Retargeting

$ python3 main.py data/images/SIGD16/4.jpg --init_from target --width_factor 1.5

Input	Output

Style transfer

$ python3 main.py data/images/style_transfer/style/mondrian.jpg --init_from data/images/style_transfer/content/trump.jpg --fine_dim 1024 --coarse_dim 256 --noise_sigma 0

Input	init_from	Output

Texture synthesis

$ python3 main.py data/images/textures/cobbles.jpeg --width_factor 1.5 --height_factor 1.5

Input	Output

Reproduce paper tables

I added the Places50 and SIGD16 datasets from Drop-The-Gan and SinGAN so that results can be reproduced

Apart from the datasets from the paper I collected some interesting retargeting images in the images folder

In the images folder you can find images I collected from various repos and papers cited in my paper.

Cite

@inproceedings{elnekave2022generating,
  title={Generating natural images with direct Patch Distributions Matching},
  author={Elnekave, Ariel and Weiss, Yair},
  booktitle={Computer Vision--ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23--27, 2022, Proceedings, Part XVII},
  pages={544--560},
  year={2022},
  organization={Springer}
}

TimZhang001/GPDM