ParaFlow: Parallel Sampling for Flow Matching Models

ICLR 2026 Conference Submission4075 Authors

Keywords: Flow Matching Models, Parallel Sampling

Abstract: Flow Matching (FM) models, including state-of-the-art architectures like Stable Diffusion 3 and Flux, have achieved remarkable success in high-fidelity data generation. However, their inference process is fundamentally autoregressive, requiring hundreds of sequential neural network evaluations to solve an ordinary differential equation (ODE), which results in slow sampling speeds. This paper challenges the inherently sequential nature of this process. We introduce \textbf{ParaFlow}, a novel parallel sampling framework that recasts the autoregressive sampling process, typically handled by numerical ODE solvers, as a system of triangular nonlinear equations (TNEs). This reformulation decouples the dependencies between sampling steps, enabling parallel computation of the vector field across multiple timesteps simultaneously. We prove that this TNE system possesses a unique solution that precisely matches the trajectory of the original autoregressive sampler. By solving this system with an efficient relaxed fixed-point iteration method, ParaFlow significantly reduces the number of required sequential operations. Extensive experiments on Stable Diffusion 3 and Flux models show that ParaFlow achieves up to a \textbf{4$\times$} reduction in sequential sampling steps and a \textbf{2-4$\times$} speedup in wall-clock time, with negligible degradation in generation quality, even at high resolutions like 2048$\times$2048.

Primary Area: generative models

Submission Number: 4075