In this work they take a different approach to the GAN model [1406.2661]. In the traditionally GAN model a neural network is trained to up-sample from random noise in a feed forward fashion to generate samples from the data distribution.
This work instead iteratively permutes an image of random noise similar to Artistic Style Transfer [1508.06576]. The image is permuted in order to fool a set of discriminators. To obtain the set of discriminators each is trained starting from random noise until some max $t$ step.
- At first a discriminator is trained to discriminate between the true data and random noise .
- Images is then permuted using gradients which aim to fool the discriminator and included in the data distribution as a negative example.
- The discriminator is trained on the true data + random noise + fake data from the previous steps
The images generated at each step are shown below:

After being trained the model is able to generate a sample by iterating over each trained discriminator and applying gradient updates on from random noise. For this storing only the weights of the discriminators is required.
Poster from ICCV2017:
