Abstract: To address the problem that Enhanced Long-Range Navigation (eLoran) signals are susceptible to cross-interference and atmospheric noise during propagation, resulting in poor received signal quality and low message demodulation accuracy, this article proposes an improved conditional diffusion model for signal restoration. For the first time, the conditional diffusion probabilistic model is introduced into the enhanced Loran signal restoration task, and a conditional generation framework with an improved one-dimensional U-Net as the backbone network is constructed. To account for the temporal structural characteristics of pulse signals, a noise estimation network integrating improved residual modules and a self-attention mechanism is designed to enhance signal feature extraction capability. Additionally, the noisy signal is incorporated as conditioning information to guide the reverse sampling process, thereby achieving high-quality signal restoration. Experimental results show that, when the input signal-to-noise ratio (SNR) is −11 dB, the output SNR of the restored signal is improved by 27.61 dB, and the message demodulation accuracy reaches 97.8%. Furthermore, experiments on real-world signals further verify the effectiveness of the proposed model, with an SNR improvement of 13 dB after restoration. The proposed model demonstrates strong generalization capability, robustness, and stability, and can provide technical support for high-quality restoration of eLoran signals.