TL;DR Preliminaries Fundamentals Conditional DMs Efficient DMs Generalized DMs References

Generative AI Meets 6G and Beyond: Diffusion Models for Semantic Communications

Hai-Long Qin1
Jincheng Dai1
Guo Lu2
Shuo Shao3
Sixian Wang2
Tongda Xu4
Wenjun Zhang2
Ping Zhang1
Khaled B. Letaief5
1Beijing University of Posts and Telecommunications (BUPT) 2Shanghai Jiao Tong University (SJTU) 3East China Normal University (ECNU) 4Tsinghua University (THU) 5Hong Kong University of Science and Technology (HKUST)
Accepted by IEEE Communications Surveys & Tutorials (COMST)
arXiv Paper GitHub
TL;DR

Preliminaries

Discriminative vs. Generative Modeling

Fundamentals of Diffusion Models

Score Matching & Langevin Dynamics

Langevin Dynamics

Score-based Modeling Pipeline
Interactive: Score Field & Langevin Dynamics click to interact

Score-Based Modeling with SDEs

Forward-Reverse SDE Pipeline
Interactive: Forward Diffusion Process drag slider

Probability Flow ODEs & Solvers

Predictor-Corrector Method
Interactive: Reverse SDE vs Probability Flow ODE animated

Conditional Diffusion Models

Inference-time Conditional Diffusion Models

Inference-time Conditioning

Classifier Guidance (CG)

Estimator Guidance (DPS & BlindDPS)

Training-time Conditional Diffusion Models

Training-time Conditioning

Classifier-Free Guidance (CFG)

Interactive: Classifier-Free Guidance Strength drag slider
1.0

Efficient Diffusion Methods

Dimensionality Reduction

Knowledge Distillation

Consistency Models

Structure Pruning

Cache Reuse

Flow Matching

Flow Matching Mechanism
Interactive: Flow Matching vs Score-based Diffusion animated

Generalized Diffusion Models

Modality Expansion

Domain Adaptation

Schrödinger Bridges

Task Generalization

References