I'm surprised so many people are interested in this toy project.

I saw some friends on Weibo who said that the training results were not good.

Here are the results of the training set after 8800 steps.

The model can fit the training images in a short time.

This ckpt may be useful, but I am not sure. https://cloudstor.aarnet.edu.au/plus/s/YHDWgez1g3RFc6o

I guess the environment should be tensorflow==1.1...

You also need to download the pretrain weight for vgg through following links:

https://github.com/machrisaa/tensorflow-vgg

Download the training images and unzip it: https://cloudstor.aarnet.edu.au/plus/s/VWZJaWfbla3kFch

Run bash autotest.sh

Generate your lover with your photo

This project is an old project from two years ago.

We collect pictures of married couples on the Internet and pre-process the images.

The model was trained by using the code of pix2pix.

An example of training images:

You can make photos of yourself with the person you loved as a training pair~

This repo is just for entertainment.

If you have any interests, please feel free to use the code and the dataset.

I am not very familiar with tensorflow...

I have abandoned tf and started using pytorch.

This project is a toy project, and I am working on semantic segmentation, instance segmentation on 2D images and videos for my Ph.D. career. If you are interested in these topics, you can follow my github. I will release a new code on video segmentation recently.

I collect the marriage images on BaiDu image by using a spider. Here is a very old and simple code for collecting images: https://blog.csdn.net/qq_27879381/article/details/65015280#comments

Just change the keyword.

If you really want to train an effective model, you can collect more images.

CVPR2019： https://github.com/irfanICMLL/structure_knowledge_distillation