We propose a flexible and theoretically supported framework for scalable Nonnegative Matrix Factorization. The goal is to find nonnegative low-rank components directly from compressed measurements, accessing the original data only once or twice. We consider compression through →