Given two images source, destination and a mask, it is to blend destination into source in a manner that is visually appealing.

We implemented an encoder-decoder network which takes low resolution(64X64) composite image(source cropped onto destination) and generates a low resolution image(64X64) which looks more natural than the composite. Using this low resolution image and using Laplacian pyramid we tried to optimize Gaussian-Poisson Equation

i) By gradient Descent

ii) Pyramid Blending

pytorch cv2 numpy

Download the data from https://www.cse.iitb.ac.in/~charith/aligned_images.tar Create train test splits Create low resolution data by using savedata(train) in src/train.py Change hyperparameters as desired Run the train function

run the script blend.py with arguments -src source_img -dest dest_img -mask mask_img -model path_to_network_weights

GP-GAN: Towards Realistic High-Resolution Image Blending https://arxiv.org/pdf/1703.07195.pdf

Result1

Result2

Look at ImageBlending.ipynb for understanding code