ETD

• autoencoders
• denoising diffusion
• diffusion backbones
• diffusion guidance
• diffusion models
• diffusion transformers
• flow matching
• generative AI
• image generation
• latent diffusion
• latent representation learning
• score-based diffusion
• stochastic differential equations
• synthetic data
• text-to-image generation

The goal of this work is to provide the reader of an understandable guide to methods underlying
the state of the art generative models; In recent years, the quality of synthetic media (Image,
Video, Audio, etc...) generation has reached a quality that has the potential to bring about
a fundamental change in our society. Modern generative frameworks mostly rely on Diffusion
models that represent the cutting-edge methodology to generate new samples from the underlying
unknown distribution of some set of observations. Diffusion models applied to high-dimensional
data such as images, videos or audio suffers from scaling difficulties; Latent diffusion methods apply
diffusion techniques to compressed latent representations of high-dimensional data employing some
autoencoder architecture, taming the dimensional scaling problem and enabling diffusion models
to focus on perceptualy relevant details. Diffusion models can also be conditioned to generate data
depending on some user specified input (such as text) via conditional generation.Moreover, most
recent methods to efficiently implement all of those ideas will be presented. This work is aimed at
readers with a basic understanding of machine learning and probability theory who seek to have
a clearer idea of how modern generative models work and what enables them, hoping to attract
new practitioners in further investigation and experimentation in this fascinating field.