References

• Pixel Recurrent Neural Networks (2016 paper)
• Conditional Image Generation with PixelCNN Decoder (2016 paper)
• tfp.distributions.PixelCNN

Autoregressive model

• Wikipedia page

$\Huge X_{t}=c+\sum {{i=1}}^{p}\varphi{i}X_{{t-i}}+\varepsilon_{t}$

$\varphi_i$ are the parameters of the model, and $\varepsilon_t$ is the noise.

Pixel by pixel generation

$\Huge p(\mathbf{x})=\Pi^{n^2}{i=1}p(x_i|x_1,\ldots,x{i-1})$