Image-Editing-using-GAN

Developing application to edit images to have desired attributes using Generative Adversarial Networks (GANs).

The objective is to have answers to questions such as -

• How would this person look if he was wearing a hat?
• How would this place look in winter season?

Models

Deep Convolutional Generative Adversarial Network (DCGAN)

Provided implementation of DCGAN model in tensorflow. Read here

Trained an encoder model(DeepConv Encoder) on top of the DCGAN model to encode an image and then pass it to the generator to produce image which is same as that of the input image. Read here

Used the Encoder-Generator model to manipulate images to have desired attributes.

Orignal Image	Encoded Image	Male	High Cheekbones	Blond Hair	Bangs

Orignal Image	Encoded Image	Bald	5 o' Clock Shadow	Double Chin	Pale Skin

Orignal Image	Encoded Image	Smiling	Bushy Eyebrows	Mustache	Wavy Hair

Cycle Consistent Adversarial Networks (CycleGAN)

Provided implementation of CycleGAN model in tensorflow. Read here

Trained the model on Monet-Photo Database and Summer-Winter Database. (with increased training images per epoch)

Monet Paintings to Photo

Input Image	Output Image	Input Image	Output Image	Input Image	Output Image

Photo to Monet Paintings

Input Image	Output Image	Input Image	Output Image	Input Image	Output Image

Author

Abhishek Tandon/ @Tandon-A