Latent Space Imaging

Figure 1. We propose an extremely-compressed imaging paradigm called Latent Space Imaging (LSI). The optical encoder (O) projects the real signal into a compressed set of measurements. A digital encoder (D θ ) then maps this signal to the latent space (L) of a frozen generative model (G), enabling image reconstruction. The L can also be linearly projected (P ) to perform downstream tasks directly-such as facial segmentation (PS), landmark detection (PL), and attribute classification (PA)-without requiring image reconstruction or a complex new model.