{%hackmd SybccZ6XD %}

Goal
> image generation

How
> GAN learns to generate image background and foregrounds separately and recursively.

## example

generated background images
> 

foreground images
> 

foreground masks

carved foreground images
> 

carved and transformed foreground images
> 

final composite images
> 

Layered structure of image
> $x = f\odot m + b\odot (1-m)$
> $f\odot m$: foreground
> $b\odot (1-m)$: background

LAYERED RECURSIVE GAN (LR-GAN)
> $x_t = ST(m_t,a_t)\odot ST(f_t,a_t) + (1 -ST(m_t,a_t))\odot x_{t-1}$
> $ST(m_t,a_t)$: affine transformed mask
> $ST(f_t,a_t)$: affine transformed appearance
> $(1 -ST(m_t,a_t))\odot x_{t-1}$: pasting on image composed so far

architecture
>