Skip to content
Official code for the paper "InGAN: Capturing and Retargeting the DNA of a Natural Image"
Python
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
examples/fruit
figs
InGAN.py
LICENSE
README.md
SceneScripts.py
configs.py
fruits.gif
monitor_60000.png
networks.py
non_rect.py
supp_video.py
test.py
train.py
train_supp_mat.py
util.py

README.md

InGAN

Official code for the paper "InGAN: Capturing and Retargeting the DNA of a Natural Image"

Project page: http://www.wisdom.weizmann.ac.il/~vision/ingan/ (See our results and visual comparison to other methods)

Accepted ICCV'19 (Oral)

If you find our work useful in your research or publication, please cite our work:

@InProceedings{InGAN,
  author = {Assaf Shocher and Shai Bagon and Phillip Isola and Michal Irani},
  title = {InGAN: Capturing and Retargeting the "DNA" of a Natural Image},
  booktitle = {The IEEE International Conference on Computer Vision (ICCV)},
  year = {2019}
}

Usage:

Test

Quick example

First you have to download the example checkpoint file, and put it in InGAN/examples/fruit/. Will defaulty run on the fruits image, using an existing checkpoint.

python test.py

General testing

By default, when testing you get a collage of various sizes and a smooth video of the transforms. You can also choose to test specific sizes, non-rectangular transforms and more.

See configs.py, for all the options. You can either edit this file or modify configuration from command-line. Examples:

python test.py --input_image_path /path/to/some/image.png  # choose input image
python test.py --test_non_rect  # also output non rectangular transformation results
python test.py --test_vid_scale 2.0, 0.5, 2.5, 0.2  # boundary scales for output video: [max_v, min_v, max_h, min_h]

Please see configs.py for many more options

Train

Quick example

Will defaulty run on the fruits image.

python train.py

General training

See configs.py for all the options. You can either edit this file or modify configuration from command-line. Examples:

python train.py --input_image_path /path/to/some/image.png  # choose input image
python train.py --G_num_resblocks 3  # change number of residual block in the generator

Please see configs.py for many more options

monitoring

In you results folder, monitor files will be periodically created, example:

Produce complex animations by scripts:

Please see the file supp_video.py

Parallel training for many images

Please see the file train_supp_mat.py

You can’t perform that action at this time.