将单像素成像迁移到潜在空间，低采样下仍能高保真重建。

In recent years, the introduction of deep learning into the field of single-pixel imaging has garnered significant attention. However, traditional networks often operate within the pixel space. To address this, we innovatively migrate single-pixel imaging to the latent space, naming this framework LSSPI (Latent Space Single-Pixel Imaging). Within the latent space, we conduct in-depth explorations into both reconstruction and generation tasks for single-pixel imaging. Notably, this approach significantly enhances imaging capabilities even under low sampling rate conditions. Compared to conventional deep learning networks, LSSPI not only reconstructs images with higher signal-to-noise ratios (SNR) and richer details under equivalent sampling rates but also enables blind denoising and effective recovery of high-frequency information. Furthermore, by migrating single-pixel imaging to the latent space, LSSPI achieves superior advantages in terms of model parameter efficiency and reconstruction speed. Its excellent computational efficiency further positions it as an ideal solution for low-sampling single-pixel imaging applications, effectively driving the practical implementation of single-pixel imaging technology.