Stack GAN

This paper's model architecture has many components, so I thought it would be good to layout the specifics of the architecture before implementing it.

Model Architecture

1. Stage-I GAN

2. Stage-2 GAN

Stage-I GAN

Input: Text embedding of the text description $(\varphi_t)$

Conditioning Augmentation (CA)

Purpose: Create $\hat{c_0}$ vector that captures the meaning of $\varphi_t$ with variations.

Process: $\varphi_t$ → FC layer → $\mu_0, \sigma_0$ → $\mathcal{N}(\mu_0(\varphi_t),\sigma_0(\varphi_t))$ → $\hat{c_0}$ sampled from this Gaussian distribution

Output: