RC-BFM cuts generative semantic decoding latency by 10x

Generative semantic communication receivers deliver high perceptual quality but suffer from prohibitive decoding latency due to diffusion-based iterative decoding (hundreds to thousands of steps) or independent endpoint coupling in existing flow matching methods that ignore the physical source–channel link. In a new paper accepted at Globecom 2026, researchers from Tsinghua University and Imperial College London introduce RC-BFM (Realization-Coupled Bridge Flow Matching), which reformulates receiver-side recovery as a realization-coupled bridge flow matching problem under explicit bandwidth and power constraints.

RC-BFM’s key innovation is initializing the decoder from a channel-induced semantic state rather than isotropic noise, and linking training pairs via a realization-coupled entropic optimal transport (RC-OT) plan that preserves each transmission’s physical channel realization while maintaining robustness to stochastic fading. The authors also identify independent coupling as the fundamental source of a conditional train–test distribution shift, and derive an end-to-end distortion bound whose discretization error decays as O(K⁻²). Experiments on CIFAR-10 and FFHQ-64×64 over AWGN and Rayleigh fading channels show RC-BFM achieves a superior fidelity–perception trade-off, reducing decoding latency by over 10× compared to diffusion-based receivers. This work is particularly relevant for real-time wireless applications like autonomous driving and AR/VR, where low-latency high-quality semantic communication is critical.