StackGAN and HDGAN

Everyday, there are interesting new ideas about how to use GANs to solve real world problems. These two papers try to solve the problem of generating high-resolution image from text.

The above figure from HDGAN gives a good illustration on the differences between the two papers (A vs. D).

Similarity

• Both use pre-trained text embedding networks (RNN).
• Both use GAN loss.
• Both generate hierarchical-resolution images, and gradually add more details to the final high-resolution images.

Difference

• StackGAN is a two-step algorithm, it first generates low-res image and then generates high-res image conditioned on the low-res image and text embedding.
• So StackGAN needs two training steps.
• HDGAN generate hierarchical-res images during in generator.
• HDGAN has two discriminators, the traditional Pair Loss and the Image Loss.
• HDGAN is end-to-end.

the local image loss (outputting a R i ×R i probability map to classify real or fake image patches).