Attribute2Image

This is the code for ECCV 2016 paper Attribute2Image: Conditional Image Generation from Visual Attributes by Xinchen Yan, Jimei Yang, Kihyuk Sohn and Honglak Lee.

Please follow the instructions to run the code.

Requirements

Attribute2Image requires or works with

• Mac OS X or Linux
• NVIDIA GPU

Installing Dependency

• Install Torch

Data Preprocessing

• For LFW dataset, please run the script to download the pre-processed dataset

./prep_cropped_lfw.sh

• Disclaimer: Please cite the LFW paper if you download this pre-processed version.

• For CelebA dataset, please download the original dataset and then run the script for pre-processing

./prep_cropped_celeba.sh

• Alternatively, you can download the pre-processed .t7 files with the following script:

./download_preprocessed_celeba.sh

• Disclaimer: Please cite the CelebA paper if you download the pre-processed .t7 files.

• For CUB dataset, please run the script to download the pre-processed dataset

./prep_cropped_cub.sh

Training (vanilla CVAE)

• If you want to train the LFW image generator, please run the script (less than 3 hours on a single Titan X GPU)

./demo_lfw_trainCVAE.sh

• If you want to train the CelebA image generator, please run the script (around 24 hours on a single Titan X GPU)

./demo_celeba_trainCVAE.sh

Training (disentangling CVAE)

• If you want to train the LFW layered image generator, please run the script (less than 5 hours on a single Titan X GPU)

./demo_lfw_trainDisCVAE.sh

• If you want to train the CUB layered image generator, please run the script (less than 3 hours on a single Titan X GPU)

./demo_cub_trainDisCVAE.sh

Visualization using Pre-trained Models

TBD

Citation

If you find this useful, please cite our work as follows:

@article{yan2015attribute2image,
  title={Attribute2Image: Conditional Image Generation from Visual Attributes},
  author={Yan, Xinchen and Yang, Jimei and Sohn, Kihyuk and Lee, Honglak},
  journal={arXiv preprint arXiv:1512.00570},
  year={2015}
}

Please contact "skywalkeryxc@gmail.com" if any questions.