UniDB: A Unified Diffusion Bridge Framework via Stochastic Optimal Control
Kaizhen Zhu, Mokai Pan, Yuexin Ma, Yanwei Fu, Jingyi Yu, Jingya Wang, Ye Shi · 2025
Read UniDB: A Unified Diffusion Bridge Framework via Stochastic Optimal Control by Kaizhen Zhu, Mokai Pan, Yuexin Ma, Yanwei Fu, Jingyi Yu, Jingya Wang, Ye Shi (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-05749.txt.
Introduction
Kaizhen Zhu and six co-authors unify diffusion bridge models, which morph one image kind into another with fixed ends, under a control theory frame. Past bridges use a classic trick to lock ends, but they oft blur fine detail and lack a full theory for why.
Their frame sets the morph as a cost trade between control effort and end miss. Past bridges fall out as the case where end miss is banned at all cost. A tunable end weight restores detail and lifts quality. The frame plugs into past bridges with tiny code changes. Tests across restore tasks confirm gains.
Catalog record
| CREATOR | Kaizhen Zhu, Mokai Pan, Yuexin Ma, Yanwei Fu, Jingyi Yu, Jingya Wang, Ye Shi |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 84 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2502-05749
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
UniDB: A Unified Diffusion Bridge Framework via Stochastic Optimal Control Kaizhen Zhu, Mokai Pan, Yuexin Ma, Yanwei Fu, Jingyi Yu, Jingya Wang, Ye Shi Abstract Recent advances in diffusion bridge models leverage Doob’s h h -transform to establish fixed endpoints between distributions, demonstrating promising results in image translation and restoration tasks. However, these approaches frequently produce blurred or excessively smoothed image details and lack a comprehensive theoretical foundation to explain these shortcomings. To address these limitations, we propose UniDB, a unified framework for diffusion bridges based on Stochastic Optimal Control (SOC). UniDB formulates the problem through an SOC-based optimization and derives a closed-form solution for the optimal controller, thereby unifying and generalizing existing diffusion bridge models. We demonstrate that existing diffusion bridges employing Doob’s h h -transform constitute a special case of our framework, emerging when the terminal penalty coefficient in the SOC cost function tends to infinity. By incorporating a tunable terminal penalty coefficient, UniDB achieves an optimal balance between control costs and terminal penalties, substantially improving detail preservation and output quality. Notably, UniDB seamlessly integrates with existing diffusion bridge models, requiring only minimal code modifications. Extensive experiments across diverse image restoration tasks validate the superiority and adaptability of the proposed framework. Our code is available at https://github.com/UniDB-SOC/UniDB/ . † † affiliationnotice: Equal contribution 1 Introduction Figure 1: Recent advances in diffusion bridge models leverage Doob’s h -transform to establish fixed endpoints between distributions, which introduces an h function into the forward process of a standard stochastic differential equation (SDE) to forcibly match the two endpoints. However, as shown in the figure, this method can lead to local blurring and distortion in the generated images. UniDB formulates the forward process as a stochastic optimal control problem and employs the penalty coefficient \gamma , balancing realistic SDE trajectories and target endpoint matching, to produce images with more realistic details. We also find that Doob’s h -transform is a special case in our framework when \gamma\rightarrow\infty . Therefore, our framework can seamlessly integrate with existing diffusion bridge models (with Doob’s h -transform). The diffusion model has been extensively utilized across a range of applications, including image generation and editing ( Ho et al., 2020 ; Kawar et al., 2022 ; Song et al., 2020 ; Xia et al., 2023 ; Li et al., 2023 ) , imitation learning ( Wu et al., 2024 ; Chi et al., 2023 ; Ze et al., 2024 ) and reinforcement learning ( Yang et al., 2023 ; Ding et al., 2024a ) , etc. Despite its versatility, the standard diffusion model faces limitations in transitioning between arbitrary distributions due to its inherent assumption of a Gaussian noise prior. To overcome this problem, diffusion models ( Dhariwal & Nichol, 2021 ; Ho & Salimans, 2022 ; Murata et al., 2023 ; Ding et al., 2024b ; Chung et al., 2022 ; Tang et al., 2024 ) often rely on meticulously designed conditioning mechanisms and classifier/loss guidance to facilitate conditional sampling and ensure output alignment with a target distribution. However, these methods can be cumbersome and may introduce manifold deviations during the sampling process. Meanwhile, Diffusion Schrödinger Bridge ( Shi et al., 2024 ; De Bortoli et al., 2021 ; Somnath et al., 2023 ) involves constraints that hinder direct optimization of the KL divergence, resulting in slow convergence and limited model fitting capability. To address this challenge, DDBMs ( Zheng et al., 2024 ) proposed a diffusion bridge model using Doob’s h -transform. This framework is specifically designed to establish fixed endpoints between two distinct distributions by learning the score function of the diffusion bridge from data, and then solving the stochastic differential equation (SDE) based on these learned scores to transition from one endpoint distribution to another. However, the forward SDE in DDBMs lacks the mean information of the terminal distribution, which restricts the quality of the generated images, particularly in image restoration tasks. Subsequently, GOUB ( Yue et al., 2023 ) extends this framework by integrating Doob’s h -transform with a mean-reverting SDE, achieving better results compared to DDBMs. Despite the promising results in diffusion bridge with Doob’s h -transform, two fundamental challenges persist: 1) the theoretical mechanisms by which Doob’s h -transform governs the bridging process remain poorly understood, lacking a rigorous framework to unify its empirical success; and 2) while effective for global distribution alignment, existing methods frequently degrade high-frequency details—such as sharp edges and fine textures—resulting in outputs with blurred or oversmoothed artifacts that compromise perceptual fidelity. These limitations underscore the need for both theoretical grounding and enhanced detail preservation in diffusion bridges. In this paper, we revisit the diffusion bridges through the lens of stochastic optimal control (SOC) by introducing a novel framework called UniDB, which formulates an optimization problem based on SOC principles to implement diffusion bridges. It enables the derivation of a closed-form solution for the optimal controller, along with the corresponding training objective for the diffusion bridge. UniDB identifies Doob’s h -transform as a special case when the terminal penalty coefficient in the SOC cost function approaches infinity. This explains why Doob’s h -transform may result in suboptimal solutions with blurred or distorted details. To address this limitation, UniDB utilizes the penalty coefficient in SOC to adjust the expressiveness of the image details and enhance the authenticity of the generated outputs. Our main contributions are as follows: • We introduce UniDB, a novel unified diffusion bridge framework based on stochastic optimal control. This framework generalizes existing diffusion bridge models like DDBMs and GOUB, offering a comprehensive understanding and extension of Doob’s h -transform by incorporating general forward SDE forms. • We derive closed-form solutions for the SOC problem, demonstrating that Doob’s h -transform is merely a special case within UniDB when the terminal penalty coefficient in the SOC cost function approaches infinity. This insight reveals inherent limitations in the existing diffusion bridge approaches, which UniDB overcomes. Notably, the improvement of UniDB requires minimal code modification, ensuring easy implementation. • UniDB achieves state-of-the-art results in various image restoration tasks, including super-resolution (DIV2K), inpainting (CelebA-HQ), and deraining (Rain100H), which highlights the framework’s superior image quality and adaptability across diverse scenarios. 2 Related Work Diffusion with Guidance. This technique tackles conditional generative tasks by leveraging a differentiable loss function for guidance without the need for additional training ( Chung et al., 2022 ; Shenoy et al., 2024 ; Bradley & Nakkiran, 2024 ) . However, it often yields suboptimal image quality and a prolonged sampling process due to the necessity of small step sizes. Most importantly, the sampling process is prone to manifold deviations and detail losses ( Yang et al., 2024 ) . Furthermore, enhancing the guidance of the diffusion typically requires the introduction of additional modules, thereby increasing the model’s computational complexity. Diffusion Schrödinger Bridge. This approach aims to determine a stochastic process, \pi^{*} that facilitates probabilistic transport between a given initial distribution P_{\text{prior}} , and a terminal distribution P_{\text{data}} ( Liu et al., 2023 ; Shi et al., 2024 ; De Bortoli et al., 2021 ) while minimizing the Kullback-Leibler (KL) divergence. However, its training process is usually intricate, involving constraints that hinder direct optimization of the KL divergence, resulting in slow convergence and limited model fitting