For a pretrained model, you can download it here:

Run the following command to translate the demo inputs.

python inference/image_translation_infer.py \
--model_path="/PATH/TO/MODEL/256/"
--image_hw=256
--input_tensor_name="sources_ph"
--output_tensor_name="custom_generated_t_style_source:0"
--input_image_path="./demo/inference_input/"
--output_image_path="./demo/inference_output/"

The input_image_path can be either one single image or a path containing images.

Download CelebA and the Getchu dataset by following the datasets guide. Then train your model using script from the training guide.

An English blog and a Chinese 中文 blog are published in early April 2018 and are available for readers with less technical background.

Please refer to technical report for details on the network structure and losses. The report is still work in progress.

Extra materials:

Presentation Slides at Anime Expo 2018

Our idea of using adaptive normalization parameters for image translation is not unique. To the best of our knowledge, at least two more work have similar ideas: MUNIT and EG-UNIT. Our model is developed around the same time period as these models.

Some key differences between our model and the two mentioned are -- we find UNet to be extremely helpful in maintaining semantic correspondence across domain, and we found that sharing all convolution filter weights speeds up training while maintaining the same output quality.

A lot of the code are adapted from online. Here is a non-exhaustive list of the repos where I borrowed code from extensively.

TF-Slim image models library

PGGAN

This personal project is developed and open sourced when I am working for Google, therefore you see Copyright 2018 Google LLC in each file. This is not an officially supported Google product. See License and Contributing for more details.