Images can be compressed to roughly 10 floating-point numbers. From a compressive sensing point of view, yes—in a lossy fashion. Under realistic assumptions on natural image structure, image size becomes largely irrelevant, collapsing representation to a handful of parameters. We’ll dive into a problem that is NP-hard in theory, yet appears effectively tractable for natural images, running in milliseconds in practice. Such compression techniques are crucial to modern research imaging systems, including diffuse optical tomography and single-pixel microscopy used in medical and biological research, defying conventional limits in signal processing, computational complexity, and conventional imaging. We’ll explore how to achieve such massive compression in Python, using quantisation, optimisation, and machine learning.