Abstract:Denoising diffusion models have recently emerged as a powerful class of generative models. They provide state-of-the-art results, not only for unconditional simulation, but also when used to solve conditional simulation problems arising in a wide range of inverse problems. A limitation of these models is that they are computationally intensive at generation time as they require simulating a diffusion process over a long time horizon. When performing unconditional simulation, a Schrödinger bridge formulation of generative modeling leads to a theoretically grounded algorithm shortening generation time which is complementary to other proposed acceleration techniques. We extend the Schrödinger bridge framework to conditional simulation. We demonstrate this novel methodology on various applications including image super-resolution, optimal filtering for state-space models and the refinement of pre-trained networks. Our code can be found at <a class="link-external link-https" href="https://github.com/vdeborto/cdsb" rel="external noopener nofollow">this https URL</a>.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is: how to perform conditional simulation efficiently in generative models, especially for tasks that require sampling from the posterior distribution \(p(x|y_{\text{obs}}) \propto p_{\text{data}}(x)g(y_{\text{obs}}|x)\). Specifically, the paper proposes a new method - Conditional Diffusion Schrödinger Bridge (CDSB) - to accelerate and improve the performance of conditional generative models.
### Problem Background
Traditional denoising diffusion models have achieved remarkable success in unconditional generation tasks, but in conditional generation tasks, these models have the problem of high computational cost. In order to sample from the posterior distribution, a long - running diffusion process usually needs to be run, which leads to high computational complexity.
### Main Contributions of the Paper
1. **Developing Conditional Schrödinger Bridge (CSB)**:
- Proposes a new Schrödinger bridge formulation method for conditional simulation.
- By extending the Diffusion SB algorithm, proposes an iterative algorithm CDSB to approximately solve the CSB problem.
2. **Improving Sampling Efficiency**:
- Uses conditional reference measure to accelerate the sampling process. For example, in super - resolution image generation, uses the low - resolution image as an approximation of the initialization noise.
- Proposes a conditional forward process, making the dynamics of the initial forward noise process depend on the observed value \(y\), thus accelerating the convergence of IPF (Iterative Proportional Fitting).
3. **Forward - Backward Sampling Scheme**:
- Explores a forward - backward sampling scheme (CDSB - FB), first sampling from the joint distribution \(p_{\text{join}}\), and then sampling through the forward half - bridge and the backward half - bridge to reduce errors.
### Application Examples
The paper demonstrates the effectiveness of CDSB in multiple applications, including:
- **Image Super - Resolution**: By using the low - resolution image as the initialization noise, improves the quality of generating high - resolution images.
- **Optimal Filtering**: Applied to the optimal filtering problem in state - space models.
- **Optimization of Pretrained Networks**: Further optimizes the pretrained super - resolution model through CDSB, making the generated samples closer to the real data distribution.
### Summary
The core problem of the paper is to improve the efficiency and accuracy of conditional generative models, especially in tasks that require sampling from the posterior distribution. By introducing the conditional Schrödinger bridge and a series of improvement measures, the method proposed in the paper not only accelerates the sampling process but also improves the quality of the generated samples.