Diffusion-Based Semantic Communications with Scene Graph Enhancement

Abstract

Semantic communication improves transmission efficiency under resource constraints by transmitting semantic information rather than raw data. However, under poor channel conditions, a recovered transmitted image is often blurred and loses fine texture. Although generative models offer a promising solution for semantic recovery, they typically rely on latent representations that implicitly model semantic structures. Consequently, when transmitted latent features are distorted by the channel, the reconstruction process becomes prone to semantic drift, particularly in complex scenes where fine-grained details and object relationships are difficult to preserve. To address these challenges, we propose a diffusion-based semantic communication framework that leverages the powerful generative capabilities of diffusion models to produce visually clearer images. Moreover, scene graph information is introduced as an additional structural semantic constraint to enhance the reconstructed image from diffusion, thereby mitigating semantic drift in generated results under channel degradation and preserving finer-grained image details in complex scenes. Experiments on image datasets validate the effectiveness of the proposed generative semantic communication framework.