Training-Free Multi-User Generative Semantic Communications via Null-Space Diffusion Sampling

Recent advances in artificial intelligence (AI) models, such as large language models and diffusion models, have shown significant potential in semantic communication by reconstructing multimedia data from highly compressed semantic signals under limited bandwidth and poor channel conditions. Unlike most existing approaches that focus on single-user scenarios with typical encoder-decoder models, this paper rethinks multi-user semantic communications using large generative models. In particular, in multi-user orthogonal frequency division multiple access (OFDMA) systems, we propose to reduce the number of subcarriers assigned per user by leveraging generative diffusion models to locally reconstruct missing or noisy information. By utilizing the null-space decomposition method for diffusion model sampling, we provide a traning-free, closed-form receiver design guideline for diffusion noise scheduling. Simulation results demonstrate that our proposed method achieves high-fidelity image reconstruction using only 60% of the original subcarriers, and outperforms deep joint source-channel coding (DeepJSCC) and LDPC baselines by up to -10 dB in signal-to-noise ratio (SNR).