/Image-Editing-using-GAN

Developing an image editing application to edit images to have a desired attribute using Generative Adversarial Networks(DCGAN and CycleGAN) (https://tandon-a.github.io/Image-Editing-using-GAN/)

Primary LanguagePythonMIT LicenseMIT

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?

Img 1

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