Multi-Scale Conditional Diffusion Model for Full-Waveform LiDAR Super-Resolution

Full-waveform LiDAR records complete temporal backscattered signals and therefore contains rich structural information about targets. However, practical received waveforms are often degraded by noise and limited temporal resolution, making accurate waveform reconstruction challenging. This paper proposes a multi-scale conditional diffusion framework for full-waveform LiDAR super-resolution. The framework adapts conditional diffusion modeling to one-dimensional LiDAR waveforms by injecting low-resolution waveform priors into a conditional 1D U-Net through multi-scale guidance and FiLM-based modulation. A late-stage high-frequency damping strategy and a hybrid structural loss are further introduced to improve reconstruction stability, preserve spectral consistency, and suppress spurious oscillatory peaks. Experiments on a simulated benchmark show that the proposed method outperforms representative baselines in L1, MSE, spectral loss, peak position error, and spurious peak rate. Additional tests on measured pulsed LiDAR waveforms further demonstrate its cross-domain generalization ability without fine-tuning.