DriftDecode: One-Step Wireless Image Decoding via Drifting-Inspired Detail Recovery

Researchers from KTH Royal Institute of Technology have introduced DriftDecode, a one-step generative decoder for wireless image transmission that dramatically reduces latency while maintaining or improving reconstruction quality. Traditional diffusion-based and flow-based decoders rely on iterative inference, which introduces substantial delay. DriftDecode instead frames wireless image decoding as a recovery task, leveraging the fact that received signals preserve the coarse structure of the source image. The decoder couples a one-step U-Net with a drift-inspired instance-level texture loss, which reformulates the drifting-field mechanism from generative drifting models in perceptual feature space. This guides each reconstructed local feature toward its spatially aligned ground-truth counterpart while suppressing mismatched textures.

In experiments on DIV2K and MNIST under additive white Gaussian noise (AWGN) and Rayleigh fading channels, DriftDecode achieved just 30ms decoding latency, a 4.8× speedup over a 10-step flow-matching decoder. It consistently outperformed standard MSE-only training, yielding up to 1.13 dB PSNR gain on MNIST under Rayleigh fading. The method demonstrates that recovery-oriented one-step decoding can be an effective alternative to iterative generative decoding for low-latency wireless image transmission, striking a favorable quality-latency tradeoff.