The paper introduces the Bidirectional Consistency Model (BCM), which enhances existing consistency models by enabling bidirectional traversal along the Probability Flow (PF) Ordinary Differential Equation (ODE). This allows the model to both generate and invert images efficiently within a single framework. The key contributions of BCM are:
1. **Unified Framework**: BCM unifies the generation and inversion tasks by learning a single neural network that can traverse the PF ODE in both forward and backward directions.
2. **Efficient Training**: BCM can be trained from scratch or fine-tuned using a pre-trained consistency model, reducing training costs and increasing scalability.
3. **Enhanced Sampling**: BCM supports multi-step sampling schemes, including ancestral sampling and zigzag sampling, which improve sample quality.
4. **Downstream Applications**: BCM demonstrates superior performance in various downstream tasks such as image interpolation, inpainting, and blind restoration of compressed images.
The paper also discusses the technical details of BCM, including its network parameterization, training objective, and sampling algorithms. Experimental results show that BCM achieves comparable or better results compared to other methods, with significantly fewer function evaluations (NFEs). Additionally, BCM outperforms other models in inversion tasks, achieving lower reconstruction errors with fewer NFEs. The authors highlight the potential of BCM in enabling broader applications and future improvements.The paper introduces the Bidirectional Consistency Model (BCM), which enhances existing consistency models by enabling bidirectional traversal along the Probability Flow (PF) Ordinary Differential Equation (ODE). This allows the model to both generate and invert images efficiently within a single framework. The key contributions of BCM are:
1. **Unified Framework**: BCM unifies the generation and inversion tasks by learning a single neural network that can traverse the PF ODE in both forward and backward directions.
2. **Efficient Training**: BCM can be trained from scratch or fine-tuned using a pre-trained consistency model, reducing training costs and increasing scalability.
3. **Enhanced Sampling**: BCM supports multi-step sampling schemes, including ancestral sampling and zigzag sampling, which improve sample quality.
4. **Downstream Applications**: BCM demonstrates superior performance in various downstream tasks such as image interpolation, inpainting, and blind restoration of compressed images.
The paper also discusses the technical details of BCM, including its network parameterization, training objective, and sampling algorithms. Experimental results show that BCM achieves comparable or better results compared to other methods, with significantly fewer function evaluations (NFEs). Additionally, BCM outperforms other models in inversion tasks, achieving lower reconstruction errors with fewer NFEs. The authors highlight the potential of BCM in enabling broader applications and future improvements.