Home

Awesome

A-Survey-on-Generative-Diffusion-Model

AwesomeGitHub stars GitHub forks visitors

A curated list for diffusion generative models introduced by the paper--A Survey on Generative Diffusion Model

Table of Contents

<!-- - [A Summary of Methodology Details](#A-Summary-of-Methodology-Details) - [A Summary of Implementation Details](#A-Summary-of-Implementation-Details) - [A Summary of Common Graph Datasets](#A-Summary-of-Common-Graph-Datasets) - [A Summary of Open-source Codes](#A-Summary-of-Open-source-Codes) -->

0. Overview

The original idea of the diffusion probabilistic model is to recreate a specific distribution that starts with random noise.

We provided Diffusion Model.pdf, the slide that serves as a vivid explanation for our article. Here, we not only thank for the articles cited in our survey, but also thank the "Tutorial on Denoising Diffusion-based Generative Modeling: Foundations and Applications" provided by NVIDIA tutorial. Besides, there's also two GitHub Repos for summarizing the up-to-date articles, Awesome-Diffusion-Models and What is the Score?. Thank you for your contribution!

<p align="center"> <img src='./figs/idea.png' width="600"> </p>

1. Methodology Improvement

<!-- ### Variational Gap Optimization --> <!-- ### Dimension Deduction -->

Nowadays, the main concern of the diffusion model is to speed up its speed and reduce the cost of computing. In general cases, it takes thousands of steps for diffusion models to generate a high-quality sample. Mainly focusing on improving sampling speed, many works from different aspects come into reality. Besides, other problems such as variational gap optimization, distribution diversification, and dimension reduction are also attracting extensive research interests. We divide the improved algorithm w.r.t. problems to be solved. For each problem, we present detailed classification of solutions.

<p align="center"> <img src='./figs/method.png' width="1000"> </p>

1.1 Speed-up

1.1.1 Training Scheme

Knowledge DIstillation

Diffusion Scheme Learning

Noise Scale Designing

Data Distribution Replace

1.1.2 Training-Free Sampling

Analytical Method

Implicit Sampler

Differential Equation Solver Sampler

Dynamic Programming Adjustment

<!-- TODO: 此前的引用, 结合自正文段落和图, 接下来的应用仅仅是图上的引用, 可能忽略段落中的部分内容. -->

1.1.3 Mixed-Modeling

Acceleration Mixture

Expressiveness Mixture

1.1.4 Score-Diffusion Unification

1.2 Distribution Diversification

1.2.1 Continuous Space

1.2.2 Discrete Space

1.2.3 Constrained Space

1.3 Likelihood Optimization

1.3.1 Improved ELBO

Score Connection

Redesign

1.3.2 Variational Gap Optimization

1.4 Dimension Reduction

1.4.1 Mixed-Modeling

<p align="center"> <img src='./figs/table1.jpg' width="900"> </p>

2. Application

Benefiting from the powerful ability to generate realistic samples, diffusion models have been widely used in various fields such as computer vision, natural language processing, and bioinformatics.

<p align="center"> <img src='./figs/v9app+.png' width="1000"> </p>

2.1 Computer Vision

2.1.1 Low-level Vision

2.1.2 High-level Vision

2.1.3 3D Vision

2.1.4 Video Modeling

2.1.5 Medical Application

2.2 Sequential Modeling

2.2.1 Natural Language Processing

2.2.2 Time Series

2.3 Audio

2.3.1 Sound Generation

2.3.2 Text to Speech

2.4 AI for Science

2.4.1 Molecular Conformation Generation

2.4.2 Material Design

<p align="center"> <img src='./figs/table2.jpg' width="1000"> </p>

Contribute

If you would like to help contribute this list, please feel free to contact me or add pull request with the following Markdown format:

- Paper Name. 
  - Author List. *Conference Year*. [[pdf]](link) [[code]](link)

This is a Github Summary of our Survey. If you find this file useful in your research, please consider citing:

@article{cao2022survey,
  title={A Survey on Generative Diffusion Model},
  author={Cao, Hanqun and Tan, Cheng and Gao, Zhangyang and Chen, Guangyong and Heng, Pheng-Ann and Li, Stan Z},
  journal={arXiv preprint arXiv:2209.02646},
  year={2022}
}

Feedback

If you have any issue about this work, please feel free to contact me by email: