Photographic Image Synthesis with Cascaded Refinement Networks

This is a Tensorflow implementation of cascaded refinement networks to synthesize photographic images from semantic layouts.

Setup

Requirement

Required python libraries: Tensorflow (>=1.0) + Scipy + Numpy + Pillow.

Tested in Ubuntu + Intel i7 CPU + Nvidia Titan X (Pascal) with Cuda (>=8.0) and CuDNN (>=5.0). CPU mode should also work with minor changes.

Quick Start (Testing)

1. Clone this repository.
2. Download the pretrained models from Google Drive by running "python download_models.py". It takes several minutes to download all the models.
3. Run "python demo_512p.py" or "python demo_1024p.py" (requires large GPU memory) to synthesize images.
4. The synthesized images are saved in "result_512p/final" or "result_1024p/final".

Training

To train a model at 256p resolution, please set "is_training=True" and change the file paths for training and test sets accordingly in "demo_256p.py". Then run "demo_256p.py".

To train a model at 512p resolution, we fine-tune the pretrained model at 256p using "demo_512p.py". Also change "is_training=True" and file paths accordingly.

To train a model at 1024p resolution, we fine-tune the pretrained model at 512p using "demo_1024p.py". Also change "is_training=True" and file paths accordingly.

Video

https://youtu.be/0fhUJT21-bs

Citation

If you use our code for research, please cite our paper:

Qifeng Chen and Vladlen Koltun. Photographic Image Synthesis with Cascaded Refinement Networks. In ICCV 2017.

Todo List

1. The MTurk scripts for evaluation.
2. Add the code and models for the GTA dataset.
3. ...

License

MIT License