PeRFlow is a flow-based method for accelerating diffusion models by dividing the sampling process into time windows and straightening the trajectories within each interval through reflow operations. This approach allows PeRFlow to achieve superior performance in few-step generation and inherit knowledge from pretrained diffusion models through dedicated parameterizations. The training converges quickly, and the resulting models demonstrate strong transferability, making them universal plug-and-play accelerators compatible with various workflows based on pretrained diffusion models. PeRFlow avoids the need for extensive synthetic data generation by using real training data and a divide-and-conquer strategy, significantly reducing computational costs. It also supports classifier-free guidance, enabling better text-image alignment and generation diversity. PeRFlow has been tested on several large-scale text-to-image and text-to-video models, including Stable Diffusion (SD) 1.5, SD 2.1, SDXL, and AnimateDiff, showing improved performance in terms of FID values, visual quality, and generation diversity. PeRFlow can be seamlessly integrated into various workflows, such as ControlNet, IP-Adaptor, and multiview generation, enhancing their performance with minimal additional steps. The method is efficient, scalable, and suitable for different data modalities. PeRFlow provides a simple and effective framework for training few-step generative flows, achieving minimal gap with diffusion models and enabling fast convergence. It is a lightweight acceleration framework that can be easily applied to training unconditional/conditional generative models of different data modalities.PeRFlow is a flow-based method for accelerating diffusion models by dividing the sampling process into time windows and straightening the trajectories within each interval through reflow operations. This approach allows PeRFlow to achieve superior performance in few-step generation and inherit knowledge from pretrained diffusion models through dedicated parameterizations. The training converges quickly, and the resulting models demonstrate strong transferability, making them universal plug-and-play accelerators compatible with various workflows based on pretrained diffusion models. PeRFlow avoids the need for extensive synthetic data generation by using real training data and a divide-and-conquer strategy, significantly reducing computational costs. It also supports classifier-free guidance, enabling better text-image alignment and generation diversity. PeRFlow has been tested on several large-scale text-to-image and text-to-video models, including Stable Diffusion (SD) 1.5, SD 2.1, SDXL, and AnimateDiff, showing improved performance in terms of FID values, visual quality, and generation diversity. PeRFlow can be seamlessly integrated into various workflows, such as ControlNet, IP-Adaptor, and multiview generation, enhancing their performance with minimal additional steps. The method is efficient, scalable, and suitable for different data modalities. PeRFlow provides a simple and effective framework for training few-step generative flows, achieving minimal gap with diffusion models and enabling fast convergence. It is a lightweight acceleration framework that can be easily applied to training unconditional/conditional generative models of different data modalities.