LoRaFlow reconstructs the raw LoRa waveform itself at extremely low SNR by treating denoising as a rectified-flow generative process, mapping the noisy received signal along a near-straight ODE trajectory back to the clean signal, and inserts this as a drop-in preprocessing stage before the standard dechirp. It therefore recovers (rather than merely classifies) symbols while staying fully compatible with existing LoRa hardware and software. The signal-recovery framing, the SNR-to-time mapping that skips ODE steps in proportion to input noise, and the hybrid Diffusion-Transformer/convolutional denoiser distinguish it from prior classification-based neural demodulators such as NELoRa.
LoRa technology, crucial for low-power wide-area networks, faces significant performance degradation at extremely low signal-to-noise ratios (SNRs). We present LoRaFlow, a novel approach using rectif…