✳flâneur — a map of the web's best reading

Diffusion models from scratch

chenyang.co · 3,769 words · saved by 3 readers

Diffusion models have recently produced impressive results in generative modeling, in particular sampling from multimodal distributions. Not only has diffusion models seen widespread adoption in text-to-image generation tools such as Stable Diffusion, they also excel in other application domains such as audio/video/3D generation, protein design, robotics path planning, all of which require sampling from multimodal distributions. This tutorial aims to introduce diffusion models from an optimization perspective as introduced in our paper. It will go over both theory and code, using the theory to explain how to implement diffusion models from scratch. By the end of the tutorial, you will learn how to implement training and sampling code for a toy dataset, which will also work for larger datasets and models. In this tutorial we will mainly reference code from smalldiffusion. For pedagogical purposes, the code presented here will be simplified from the original library code, which is on its

Diffusion models from scratch, from a new theoretical perspective \[\newcommand{\Kset}{\mathcal{K}} \newcommand{\distK}{ {\rm dist}_{\Kset} } \newcommand{\projK}{ {\rm proj}_{\Kset} } \newcommand{\eps}{\epsilon} \newcommand{\Loss}{\mathcal{L}} \newcommand{\norm}[1]{\left\lVert #1 \right\lVert} \newcommand{\R}{\mathbb{R}} \DeclareMathOperator{\softmin}{softmin} \DeclareMathOperator{\distop}{dist}\] Paper (ICML 2024) Code (Github) Discussion on HackerNews Diffusion models have recently produced impressive results in generative modeling, in particular sampling from multimodal distributions. Not o

Explore this link on the map →

saved by

related reading