Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Enabling Training-Free Semantic Communication Systems with Generative Diffusion Models

Shunpu Tang, Yuanyuan Jia, Qianqian Yang, Ruichen Zhang, Jihong Park, Dusit Niyato

TL;DR

This work addresses training‑free SemCom by leveraging pretrained generative diffusion models (GDMs). It introduces a two‑stage forward diffusion scheme using DDIM inversion, splitting diffusion between transmitter and receiver, and optimizes sampling steps to compensate for channel noise, demonstrating significant gains on Kodak images with a BCR of 1/96. The key contributions include a practical encoding/decoding pipeline based on a fixed Stable Diffusion setup, analytical insights into latent distribution under channel perturbations, and extensive ablations confirming the importance of diffusion stage partitioning and denoising steps. The approach offers a scalable, training‑free alternative to task‑specific SemCom systems, with practical impact for robust image transmission in noisy wireless channels.

Abstract

Semantic communication (SemCom) has recently emerged as a promising paradigm for next-generation wireless systems. Empowered by advanced artificial intelligence (AI) technologies, SemCom has achieved significant improvements in transmission quality and efficiency. However, existing SemCom systems either rely on training over large datasets and specific channel conditions or suffer from performance degradation under channel noise when operating in a training-free manner. To address these issues, we explore the use of generative diffusion models (GDMs) as training-free SemCom systems. Specifically, we design a semantic encoding and decoding method based on the inversion and sampling process of the denoising diffusion implicit model (DDIM), which introduces a two-stage forward diffusion process, split between the transmitter and receiver to enhance robustness against channel noise. Moreover, we optimize sampling steps to compensate for the increased noise level caused by channel noise. We also conduct a brief analysis to provide insights about this design. Simulations on the Kodak dataset validate that the proposed system outperforms the existing baseline SemCom systems across various metrics.

Enabling Training-Free Semantic Communication Systems with Generative Diffusion Models

TL;DR

This work addresses training‑free SemCom by leveraging pretrained generative diffusion models (GDMs). It introduces a two‑stage forward diffusion scheme using DDIM inversion, splitting diffusion between transmitter and receiver, and optimizes sampling steps to compensate for channel noise, demonstrating significant gains on Kodak images with a BCR of 1/96. The key contributions include a practical encoding/decoding pipeline based on a fixed Stable Diffusion setup, analytical insights into latent distribution under channel perturbations, and extensive ablations confirming the importance of diffusion stage partitioning and denoising steps. The approach offers a scalable, training‑free alternative to task‑specific SemCom systems, with practical impact for robust image transmission in noisy wireless channels.

Abstract

Semantic communication (SemCom) has recently emerged as a promising paradigm for next-generation wireless systems. Empowered by advanced artificial intelligence (AI) technologies, SemCom has achieved significant improvements in transmission quality and efficiency. However, existing SemCom systems either rely on training over large datasets and specific channel conditions or suffer from performance degradation under channel noise when operating in a training-free manner. To address these issues, we explore the use of generative diffusion models (GDMs) as training-free SemCom systems. Specifically, we design a semantic encoding and decoding method based on the inversion and sampling process of the denoising diffusion implicit model (DDIM), which introduces a two-stage forward diffusion process, split between the transmitter and receiver to enhance robustness against channel noise. Moreover, we optimize sampling steps to compensate for the increased noise level caused by channel noise. We also conduct a brief analysis to provide insights about this design. Simulations on the Kodak dataset validate that the proposed system outperforms the existing baseline SemCom systems across various metrics.
Paper Structure (13 sections, 1 theorem, 17 equations, 3 figures, 2 tables, 1 algorithm)

This paper contains 13 sections, 1 theorem, 17 equations, 3 figures, 2 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. System model of SemCom
  4. Generative Diffusion Models
  5. Proposed System
  6. Semantic Encoding via DDIM Inversion
  7. Semantic Decoding via DDIM Inversion and Sampling
  8. Analysis
  9. Simulations
  10. Settings
  11. Effectiveness of the Proposed Method
  12. Ablation Studies
  13. Conclusion

Key Result

Proposition 1

Consider a latent feature $\bm{z}_0$ undergoing $T_{F,1}$ steps of the forward diffusion process, followed by normalization with a scaling factor $\gamma$. After being corrupted by the AWGN with variance $\sigma_{\rm ch}^2$, the latent further undergoes $T_{F,2}$ forward diffusion steps. The resulti where

Figures (3)

  • Figure 1: Overview of the proposed system, where the semantic encoder and decoder are all based on the pretrained stable diffusion model.
  • Figure 2: Reconstructed performance comparison for different methods, where BCR is set to 1/96 and SNR varies from 0 to 20dB
  • Figure 3: Visual comparison at SNR of 5 dB.

Theorems & Definitions (5)

  • Proposition 1
  • proof : Proof Sketch
  • Remark 1: Necessity of applying forward diffusion at the transmitter
  • Remark 2: Necessity of continuing forward diffusion at the receiver
  • Remark 3: Necessity of performing additional denoising steps