This paper introduces the Bidirectional Consistency Model (BCM), a novel approach that unifies generation and inversion tasks within a single framework. BCM learns a single neural network that enables both forward and backward traversal along the Probability Flow Ordinary Differential Equation (PF ODE), allowing for efficient generation and inversion. Unlike previous methods that require iterative evaluations of the network, BCM can generate images or invert a given image with a single Number of Function Evaluation (NFE), and can achieve improved sample quality or lower reconstruction error by chaining multiple time steps. BCM also supports more complex sampling strategies, such as ancestral sampling and zigzag sampling, which have been shown to improve sample quality. The model is capable of performing downstream tasks such as interpolation and inpainting, and has been demonstrated to achieve superior performance in these tasks. BCM can be trained from scratch or fine-tuned using a pretrained consistency model, reducing training cost and increasing scalability. The paper also discusses the limitations of the proposed method, including the plateauing of performance improvements with more NFEs and the potential for imperfect inversion. Overall, BCM provides a more efficient and effective approach to image generation and inversion compared to existing methods.This paper introduces the Bidirectional Consistency Model (BCM), a novel approach that unifies generation and inversion tasks within a single framework. BCM learns a single neural network that enables both forward and backward traversal along the Probability Flow Ordinary Differential Equation (PF ODE), allowing for efficient generation and inversion. Unlike previous methods that require iterative evaluations of the network, BCM can generate images or invert a given image with a single Number of Function Evaluation (NFE), and can achieve improved sample quality or lower reconstruction error by chaining multiple time steps. BCM also supports more complex sampling strategies, such as ancestral sampling and zigzag sampling, which have been shown to improve sample quality. The model is capable of performing downstream tasks such as interpolation and inpainting, and has been demonstrated to achieve superior performance in these tasks. BCM can be trained from scratch or fine-tuned using a pretrained consistency model, reducing training cost and increasing scalability. The paper also discusses the limitations of the proposed method, including the plateauing of performance improvements with more NFEs and the potential for imperfect inversion. Overall, BCM provides a more efficient and effective approach to image generation and inversion compared to existing methods.