TransDiff: Bridging Vision Transformers and Diffusion Models for Robust RIS-mmWave Channel Estimation

Acquiring accurate channel state information (CSI) is crucial for stable reconfigurable intelligent surface (RIS)-assisted millimeter-wave (mmWave) links; however, the passive nature of RIS elements leads to a severely high-dimensional channel. Existing diffusion-based approaches show potential, but they primarily rely on localized convolutional neural networks (CNNs) for noise prediction. While deep CNNs can theoretically model global context, they often incur substantial computational overhead to capture the long-range spatial correlations inherent in massive RIS arrays. We present TransDiff, a generative approach that integrates a lightweight Vision Transformer (ViT) into the reverse diffusion process. Rather than relying solely on standard convolutional bottlenecks, TransDiff employs self-attention mechanisms to efficiently learn global dependencies within the sparse channel structure. Numerical results demonstrate that our approach surpasses the leading CNN-based baseline (BRCNet), yielding a 3 dB improvement in normalized mean square error (NMSE) and a spectral efficiency gain of 0.8 bps/Hz, particularly in low signal-to-noise ratio (SNR) regimes.