PeRFlow is a novel method for accelerating diffusion models by dividing the sampling process into several time windows and straightening the trajectories in each interval using a reflow operation. This approach simplifies the flow trajectories and preserves the continuous flow of the original pretrained diffusion models. PeRFlow achieves superior performance in few-step generation and converges quickly due to dedicated parameterizations that inherit knowledge from pretrained diffusion models. The method is compatible with various workflows based on pre-trained diffusion models, making it a universal plug-and-play accelerator. Extensive experiments on Stable Diffusion (SD) 1.5, SD 2.1, SDXL, and AnimateDiff show that PeRFlow can generate high-quality results within four steps, outperforming other acceleration methods in terms of FID values, visual quality, and generation diversity. PeRFlow is simple, flexible, and efficient, suitable for training unconditional/conditional generative models across different data modalities.PeRFlow is a novel method for accelerating diffusion models by dividing the sampling process into several time windows and straightening the trajectories in each interval using a reflow operation. This approach simplifies the flow trajectories and preserves the continuous flow of the original pretrained diffusion models. PeRFlow achieves superior performance in few-step generation and converges quickly due to dedicated parameterizations that inherit knowledge from pretrained diffusion models. The method is compatible with various workflows based on pre-trained diffusion models, making it a universal plug-and-play accelerator. Extensive experiments on Stable Diffusion (SD) 1.5, SD 2.1, SDXL, and AnimateDiff show that PeRFlow can generate high-quality results within four steps, outperforming other acceleration methods in terms of FID values, visual quality, and generation diversity. PeRFlow is simple, flexible, and efficient, suitable for training unconditional/conditional generative models across different data modalities.