Implicit generation with pretrained priors for single-pixel imaging

Single-pixel imaging (SPI) enables image recovery from highly compressed measurements but remains sensitive to weak or mismatched priors. We present an implicit generation framework that reconstructs SPI images by optimizing the input of fixed pretrained networks under measurement consistency. By cascading denoising and super-resolution networks, the proposed method jointly exploits complementary learned priors, while an Anderson-accelerated fixed-point reset explicitly searches for prior-data-consistent solutions to improve reconstruction. Experimental results show improved detail preservation and noise suppression over state-of-the-art algorithms, enabling reconstruction at sampling ratios down to 3%.