D-SIC: Energy-Efficient Digital Semantic Image Communication via Large Generative Models

Diffusion model-based generative Large AI Models (LAMs) have shown great ability in image generation tasks. In this paper, we present a novel approach to Semantic Image Communication (SIC) by integrating Stable Cascade, a recent generative LAM, into a digital communication framework. The majority of existing Semantic Communication (SemCom) approaches operate under the assumption that the extracted information of interest, typically floating-point numbers, can be transmitted over the communication channel in an analog manner. This is a critical limitation as it makes these approaches incompatible with the prevalent digital communication architecture. Moreover, we demonstrate that such analog source-channel mappings overestimate the true performance of SemCom systems at a low signal-to-noise ratio (SNR). To ensure compatibility with existing digital systems, we introduce D-SIC, a Digital-SIC approach where the extracted semantic information is converted into bits, digitally modulated, and transmitted over the communication channel. To save bandwidth, we do not employ any channel coding and rely on the model's noise mitigation capabilities for image reconstruction. Furthermore, we analyze performance not only in terms of image quality metrics but also by considering the number of transmitted bits and total energy consumption. We show that D-SIC outperforms conventional digital image compression codecs, JPEG2000 and Better Portable Graphics (BPG), with Low-Density Parity-Check (LDPC) coding in perceptual quality metrics and energy efficiency.