Efficient Diffusion Models for Inverse Problems in Single-Photon Imaging

Single-photon avalanche diode (SPAD) detectors offer exceptional temporal resolution and sensitivity. However, due to low spatial resolution (in cheap sensors), signal sparsity, and the presence of non-Gaussian noise, particularly in low-light conditions, reconstructing images requires solving challenging inverse problems like super-resolution, denoising, and deblurring. This presentation addresses these challenges through a two-fold advancement in generative imaging. First, we demonstrate the problem of depth map super-resolution using Denoising Diffusion Probabilistic Models (DDPMs). By combining multiscale representations of Time-of-Flight (ToF) histograms with reflectivity guidance, this method provides high-resolution depth maps. Second, to address the computational bottleneck of slow diffusion sampling, we introduce a new distillation-based acceleration framework. This approach utilizes a Bayesian variational formulation to enable fast inference that remains robust to the non-Gaussian (Poisson/Binomial) noise statistics inherent in SPAD data. Currently validated on 2D photon-starved imaging, this work resolves the trade-off between reconstruction quality and speed, establishing a strong foundation for future real-time 3D depth sensing.