cheolho/generating_images_part_by_part

Implementation of the paper "Generating images part by part with composite generative adversarial networks"

Generating images part by part with composite generative adversarial networks

Tensorflow implementation of the paper "Generating images part by part with composite generative adversarial networks".

The Composite GANs (CGAN) disentangles complicated factors of images with multiple generators in which each generator generates some part of the image. Those parts are combined by an alpha blending process to create a new single image. For example, it can generate background, face, and hair sequentially with three generators. There is no supervision on what each generator should generate.

Dependencies

• tensorflow 0.10.0rc0+
• h5py 2.3.1+
• scipy 0.18.0+
• Pillow 3.1.0+

How to use

First, go into the 'data' directory and download celebA dataset.

cd data
python download.py celebA

Preprocess the celebA dataset to create a hdf5 file. It resizes images to 64*64.

python preprocess.py

Finally, go into the 'code' directory and run 'main_cgan_triple_alpha_celeba.py'.

cd ../code
python main_cgan_triple_alpha_celeba.py

You can see the samples in 'samples' directory.

References

[1] Hanock Kwak and Byoung-Tak Zhang. "Generating Images Part by Part with Composite Generative Adversarial Networks." arXiv preprint arXiv:1607.05387 (2016).