New Sampler Improves Discrete Flow Matching with Fewer Evaluations.

A new sampling method, the Time-Reparameterized Cumulative Intensity Extrapolation (TR-CIE) sampler, has been proposed to enhance Discrete Flow Matching (DFM) for generative modeling. DFM, which operates on discrete state spaces via continuous-time Markov chain dynamics, often relies on discretizations like tau-leaping for sampling. However, efficient sampling with a limited number of function evaluations (NFE) has remained a challenge. TR-CIE addresses this by incorporating two main components. First, a schedule-based time reparameterization rescales the time grid according to the noise schedule. This transformation effectively absorbs schedule-dependent growth terms and reduces stiffness, particularly near the terminal sampling stage. Second, a cumulative-intensity extrapolation updating rule is introduced. This rule reuses cached model outputs from previous steps as a history term, thereby improving the approximation of stepwise cumulative intensities on the resulting non-uniform time grid. The sampler requires only one NFE per step, incurring no additional model evaluations compared to standard tau-leaping. Theoretical analysis supports its local approximation error bounds and convergence. Extensive experiments across synthetic tasks, text generation, and text-to-image benchmarks demonstrate that TR-CIE significantly improves sampling quality when function evaluations are restricted.