Generation¶

Generative Adversarial Network (GAN)¶

GAN models are known for potentially unstable training and less diversity in generation due to their adversarial training nature.

VAE¶

VAE relies on a surrogate loss.

Flow-based Model¶

Flow models have to use specialized architectures to construct reversible transform.

Normalizing flows is a class of generative models focusing on mapping a complex probability distribution to a simple distribution such as a Gaussian. * normalizing flow的正向和逆向是完全一一对应的可逆过程，不需要在base和target distribution重复可视化 - 李宏毅：Flow-based Generative Model

Diffusion Model¶

Diffusion model：独立高斯分布可加性

overviews of different generative models

Reference¶

What are Diffusion Models?

Conditional Generation¶

consider learning a conditional mapping function $<span class="arithmatex"><span class="MathJax_Preview">G: \mathcal X \rightarrow \mathcal Y</span><script type="math/tex">G: \mathcal X \rightarrow \mathcal Y$ which generates an output $<span class="arithmatex"><span class="MathJax_Preview">\mathbf y \in \mathcal Y</span><script type="math/tex">\mathbf y \in \mathcal Y$ . Our goal is to learn a multi-modal mapping $<span class="arithmatex"><span class="MathJax_Preview">G: \mathcal X \times \mathcal Z \rightarrow \mathcal Y</span><script type="math/tex">G: \mathcal X \times \mathcal Z \rightarrow \mathcal Y$ such that an input $x$ can be mapped to multiple and diverse ouputs in $\mathcal Y$ depending on the latent factors encoded in $\mathbf z \in \mathcal Z$ .

offen suffers from mode-collapse problem.