Wireless image transmission via joint source-channel coding guided by diffusion models and multimodal semantics

Wireless image transmission based on deep joint source-channel coding (Deep JSCC) often suffers from perceptual quality degradation under low signal-to-noise ratio (SNR) conditions. To address this, this paper presents SemDiff-JSCC, a framework guided by multimodal semantics. By leveraging text descriptions and edge maps to guide the diffusion-based denoising process, the proposed approach enhances image reconstruction accuracy over noisy wireless channels. An OmniControl-based parameter reuse strategy is adopted to reduce complexity, alongside a blind channel estimation mechanism for robust adaptation to slow fading channels without dedicated pilots. Experimental results demonstrate that SemDiff-JSCC consistently outperforms baselines, particularly in scenarios with unknown SNR or channel state information. Notably, at 0 dB SNR and an extreme compression ratio of 1/128, it achieves over 21% improvement in the Fréchet inception distance (FID) metric compared with the strongest baseline at the same rate. Furthermore, it attains superior perceptual quality to existing schemes operating at a 8× higher bandwidth (R=1/16), highlighting its efficacy for high-fidelity wireless semantic communication.