Usage | Web App | | Paper | Supplementary Material | More results |

code for paper ``Image Inpainting with Contextual Reconstruction Loss". This repo (including code and models) are for research purposes only.

0. Download code and model

git clone --single-branch https://github.com/zengxianyu/crfill

download model files and put them in the ./files/ directory

1. Install dependencies:

conda env create -f environment.yml

or manually install these packages in a Python 3.6 enviroment:

pytorch=1.3.1, opencv=3.4.2, tqdm

2. Use the code:

with GPU:

python test.py --image path/to/images --mask path/to/hole/masks --output path/to/save/results

without GPU:

python test.py --image path/to/images --mask path/to/hole/masks --output path/to/save/results --nogpu

path/to/images is the path to the folder of input images; path/to/masks is the path to the folder of hole masks; path/to/save/results is where the results will be saved.

Hole masks are grayscale images where pixel values> 0 indicates the pixel at the corresponding position is missing and will be replaced with generated new content.

📣 📣 The white area of a hole mask should fully cover all pixels in the missing regions. 📣 📣

flask, requests, pillow

Then execute the following:

With GPU:

cd app
python hello.py

Without GPU:

cd app
python hello.py --nogpu

After that, open http://localhost:2334 in the browser

To use the adjusted model for high-res inpainting (specifiy the option --nogpu to run on cpu```):

python test.py --opt nearestx2 --load ./files/model_near512.pth \
--image path/to/images \
--mask path/to/hole/masks \
--output path/to/save/results \

By default, the Web APP selects from the two models based on the image size. The adjusted model will be used if the short side >= 512. To mannualy specify the model used in the Web APP:

cd app
python hello.py --opt nearestx2 --load ./files/model_near512.pth

or

cd app
python hello.py --opt convnet --load ./files/model_256.pth

The auxiliary network (used during training) is defined in networks/auxiliary.py. The full training code will be released after paper accepted.