UNIC-Lab/RadioDiff

What it solves

RadioDiff addresses the challenge of constructing radio maps (RMs)—which map pathloss based on location—without relying on computationally expensive simulations or costly physical sampling measurements. It overcomes the limitations of previous neural network methods that used discriminative modeling, which often resulted in suboptimal performance when trying to generate these complex maps.

How it works

RadioDiff treats radio map construction as a conditional generative problem using a denoised diffusion-based approach. Its architecture consists of:

• Attention U-Net with Adaptive FFT: A backbone network designed to better extract features from dynamic environments.
• Decoupled Diffusion Model: A specialized diffusion process used to further refine construction performance.
• Two-Stage Training: The system first trains an AutoEncoder (VAE) and then trains a latent diffusion-edge model to generate the final radio maps.

Who it’s for

This project is primarily for researchers and engineers working on 6G network applications, wireless communication, and radio propagation modeling who need high-accuracy pathloss estimation in dynamic environments.

Highlights

• Sampling-Free: Constructs radio maps without needing extensive physical measurements.
• State-of-the-Art Performance: Outperforms existing methods in accuracy, structural similarity, and peak signal-to-noise ratio (PSNR).
• Dynamic Environment Support: Specifically optimized for extracting features from changing environments via an adaptive fast Fourier transform module.
• Extensible Ecosystem: Part of a family of models including physics-informed (RadioDiff-k²), 3D (RadioDiff-3D), and few-shot (RadioDiff-FS) versions.