Abstract
Diffusion models (DMs) have emerged as a cornerstone of modern generative modeling, widely acclaimed for their outstanding capabilities across diverse tasks. However, their inherently slow sampling process poses substantial challenges to practical deployment. Although recent advances have accelerated sampling, these methods still suffer from cumulative errors that degrade generative quality. To remedy this, we propose a novel partial-path rectification framework that mitigates error accumulation by constructing ground-truth one-to-one image pairs. Concretely, we decompose the entire transport path into two sub-paths according to a divide-and-conquer design philosophy. Furthermore, to ensure flow consistency between the two sub-paths, we introduce a consistency training strategy that aligns the velocity field of the second sub-path with that of the first, yielding a unified and coherent sampling path. By leveraging this, our framework not only achieves high-quality one-step generation but also enables a smooth trade-off between image quality and sampling efficiency. Extensive experiments on four benchmark datasets validate the superiority of our framework.
