UNIC-Lab/RadioDiff

Implements a conditional latent diffusion model with attention U-Net backbone and adaptive FFT modules to generate high-fidelity radio maps from sparse measurements, eliminating the need for costly pathloss sampling. The two-stage training pipeline uses a variational autoencoder for dimensionality reduction followed by diffusion model training in latent space, with inference speed controllable via sampling timesteps. Integrates PyTorch and HuggingFace Accelerate for multi-GPU distributed training on the RadioMapSeer dataset, supporting dynamic environmental feature extraction through decoupled diffusion mechanisms.

306 stars.