Learning Latent Diffusion Model for Channel Modeling and Generating

Modeling the distribution of Channel State Information (CSI) and generating high quality CSIs are important for the new generation intelligent wireless communication systems, due to the demand of numerous training data for deep learning based algorithms. As the bandwidth and number of antennas increase, the increasingly complex distribution of CSIs poses significant challenges to existing deep learning-based generative models, such as Generative Adversarial Networks (GANs) and Variational AutoEncoders (VAEs). In this paper, we propose a more advanced latent diffusion model based method, dubbed Latent Diffusion Gaussian Auto-Encoder based Channel Generative Models (LDGAE) to model and generate CSIs. Specifically in LD-GAE, we design an auto-encoder with spatial noisy latent space named Spatial Gaussian Auto-Encoder (SGAE) to extract the time-space latent features and recover the CSIs by end-to-end training. We further develop a Denoising Diffusion Probabilistic Models for Latent Space (DMLS) to model the latent space of the SGAE. Simulation results show that our proposed LD-GAE can model the distribution of CSIs in single and multiple scenarios well. The generated CSIs by LD-GAE exhibit superior quality compared with Channel-GAN and vanilla VAE, and approach the real CSIs.