LoRaFlow: High-Quality Signal Reconstruction Using Rectified Flow

LoRa networks enable battery-powered IoT devices to deliver data over kilometre distances, yet their chirp-spread-spectrum links collapse once the gateway’s signal-to-noise ratio (SNR) falls below the dechirp threshold. We present LoRaFlow—a gateway-side, rectified-flow diffusion model that reconstructs the full complex LoRa waveform from heavily corrupted IQ samples, allowing reliable demodulation at SNR several decibels lower than the conventional limit. LoRaFlow combines a lightweight DiT backbone with straight-line rectified trajectories and is trained mainly on procedurally generated chirps plus a single real sample per class. Across –40 to –10 dB, it reduces symbol-error rate by up to 40% relative to standard dechirp and achieves roughly one-third fewer errors than the recent NELoRa decoder, while using one model for all spreading factors and demanding no firmware changes on end devices. By recovering packets that would otherwise be lost, LoRaFlow extends communication range, curtails retransmissions, and lengthens sensor lifetime, offering a practical drop-in upgrade for existing LoRaWAN gateways and illustrating how diffusion models can bolster other low-power wide-area networks.