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

Stage-I GAN
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:

PreviousLearning Representation For Automatic Colorization NextStyle Transfer

Last updated 2 years ago

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: