Semantic Image Inpainting with Deep Generative Models

PyTorch implementation of research paper Semantic Image Inpainting with Deep Generative Models by R.A. Yeh et al.

saved_model contains pretrained GAN parameter dictionary required for inference during inpainting
images contains image files for inpainting img contains image files for GAN training trainGAN.py script to train and save the state dictionary of GAN main.py entry point of script for inpainting

Install all dependencies

pip install -r requirements.txt

To train GAN

1. This script assumes that the path to training images provided has a subfolder, and all images are inside that subfolder. To train GAN

$ python trainGAN.py

3. Current Saved model is trained on CelebA dataset. Saved model is present in saved_model folder.

To inpaint an image

1. Script takes clean image(original image) as input and generates a patchy image out of that and tries to recover that patchy image. Meanwhile the original image is not used.

$ python main.py

After the inpainting task is completed the inpainted image is saved at the desired location

Results

• celebA Dataset