Deep Generative Image Models using a Laplacian Pyramid of Adversarial Networks
Emily Denton,Soumith Chintala, Arthur Szlam, Rob Fergus, NIPS proceedings, PP:1-9,2015
From the two previous two supportive materials it is clear what is GAN and C-GAN. This paper is a followup of the previous 2 papers, here both G and D are convolutional networks and trained using backpropagation. In case of C-GANs, the conditioning variable used here is another image which is also generated by using another C-GAN model. Whereas, in the previous material the conditioning variable was either a label or a face attribute.
It is important to know how laplacian pyramid is formed. First the gaussian pyramid is built by convolving the original image with the gaussian kernel and the resulting image is a low pass filtered image of the original version. This image is upsampled by a factor of 2 so that the sizes are compatible.
The sampling procedure is as follows: Initially, a noise model is used along with a generative model to generate a image which is then upsampled and this image is used as a conditioning image or variable along with the generative model and noise in the next level. It produces a difference image which is added with the upsampled image to create a sample. In this way the process is repeated to obtain the sample at full resolution.
The training procedure is even more interesting: Image (I0) is downsampled by a factor of 2 to produce image (I1) , this (I1) is upsampled which gives the low pass version (l) of the original image. (l) and noise is given to the generator to generate a image. The difference image , the low pass version image and image generated by using generator is given as input to the discriminator to find whether the image obtained is real/ fake. code
It is important to know how laplacian pyramid is formed. First the gaussian pyramid is built by convolving the original image with the gaussian kernel and the resulting image is a low pass filtered image of the original version. This image is upsampled by a factor of 2 so that the sizes are compatible.
The sampling procedure is as follows: Initially, a noise model is used along with a generative model to generate a image which is then upsampled and this image is used as a conditioning image or variable along with the generative model and noise in the next level. It produces a difference image which is added with the upsampled image to create a sample. In this way the process is repeated to obtain the sample at full resolution.
The training procedure is even more interesting: Image (I0) is downsampled by a factor of 2 to produce image (I1) , this (I1) is upsampled which gives the low pass version (l) of the original image. (l) and noise is given to the generator to generate a image. The difference image , the low pass version image and image generated by using generator is given as input to the discriminator to find whether the image obtained is real/ fake. code