What is the Revision Network?
The Revision Network is a style transfer module that aims to revise the rough stylized image by generating a residual details image while ensuring proper distribution of global style pattern. It also makes it easier to learn to revise local style patterns.
How does the Revision Network work?
The Revision Network follows a simple yet effective encoder-decoder architecture consisting of one down-sampling and one up-sampling layer. The model generates the final stylized image by combining the rough stylized image and the residual details image. A patch discriminator is used to capture fine patch textures under an adversarial learning setting.
What is a patch discriminator?
A patch discriminator is a type of discriminator used in the Revision Network that captures fine patch textures under an adversarial learning setting.
How is the patch discriminator defined in the Revision Network?
The patch discriminator in the Revision Network is defined following SinGAN, with five convolution layers and 32 hidden channels. The relatively shallow patch discriminator is chosen to avoid overfitting since there is only one style image, and to control the receptive field to ensure it can only capture local patterns.
What are the benefits of using the Revision Network?
Using the Revision Network helps to improve the quality of stylized images by generating a residual details image that revises local style patterns, while ensuring proper distribution of global style pattern. It is also a relatively simple yet effective model that avoids overfitting and captures local patterns through the use of a patch discriminator.
