This is a fork of https://github.com/JiahuiYu/generative_inpainting (the official code release of inpainting with contextual attention https://arxiv.org/abs/1801.07892) which attempts to implement the features described in their subsequent paper introducing Gated Convolutions (https://arxiv.org/pdf/1806.03589.pdf) and some other experimental features such as multiresolution contextual attention.
We implemented a GUI to select the watermark region of a picture before processing.
generative inpainting #1. make sure you have installed requested module
tensorflow 1.x
pillow
opencv
tkinter
neuralgym (run pip install git+https://github.com/JiahuiYu/neuralgym)
- run GUI.py
Choose picture
Circle the position of water mark
Input picture and mask
Start processing
Output processed picture
- check the processed picture in your saving path
Choose the photo you want to remove watermark->Circle the position of watermark->Get the result after generating
Sample results
original picture->our trained model->pretrained model
To train the model, I offer my dataset at Google Drive
Use gen.py to generate your dataset file, make sure it is in the same directory as your dataset folder
Modify the inpaint.yml and set the dataset path to your flist path
Please put your trained model or the pretrained model in the model_logs folder to make sure the GUI run correctly.
your project structure should look like this:
