Anime Generation
Clone or download
Latest commit cabd0d4 May 25, 2017
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
asset readme May 25, 2017
LICENSE Initial commit May 23, 2017
README.md readme May 25, 2017
data_utils.py gaussian sample May 25, 2017
improved_WGAN.py bug fix May 23, 2017
main.py add code May 23, 2017
model.py add code May 23, 2017

README.md

Conditional GAN

Conditional Generative Adversarial Networks for anime generation (AnimeGAN).

image
Training results dump every 500 min-batch in 25 epoch(26000th min-batch) for the following tags

  • blue hair blue eyes
    image
  • gray hair green eyes
    image
  • green hair red eyes
    image
  • orange hair brown eyes
    image
  • blonde hair gray eyes
    image
  • pink hair aqua eyes
    image

Sample training data

image

Environment

python3
tensorflow 1.0
scipy

Model structure

image

Data

source link
google drive link

Usage

  1. Download hw3 data from data link, place the MLDS_HW3_dataset/ in the same directory and unzip the face.zip in MLDS_HW3_dataset/
  2. Replace the tags in MLDS_HW3_dataset/sample_testing_text.txt to the right format.
  3. Start training !

Train

First time use, you need to do the preprocessing

$ python3 main.py --prepro 1

If you already have done the preprocessing

$ python3 main.py --prepro 0

Model

  • dcgan structure
  • use one hot encoding for condition tags

Test

This code will automatically dump the results for the tags specified in MLDS_HW3_dataset/sample_testing_text.txt every dump_every batches to the test_img/ folder.

Testing tags format

1,<Color> hair <Color> eyes 
2,<Color> hair <Color> eyes
3,<Color> hair <Color> eyes
4,<Color> hair <Color> eyes
.
.
.
  • Possible colors for eyes
['<UNK>', 'yellow', 'gray', 'blue', 'brown', 'red', 'green', 'purple', 'orange',
 'black', 'aqua', 'pink', 'bicolored']
  • Possible colors for hair
['<UNK>', 'gray', 'blue', 'brown', 'red', 'blonde', 'green', 'purple', 'orange',
 'black', 'aqua', 'pink', 'white']