Unifying Heterogeneous Degradations: Uncertainty-Aware Diffusion Bridge Model for All-in-One Image Restoration
UDBM introduces an uncertainty-guided diffusion bridge for single-step multi-task image restoration, outperforming state-of-the-art methods with 6× faster inference.
Key Findings
Methodology
UDBM formulates AiOIR as a pixel-wise uncertainty-driven stochastic transport problem. It employs a relaxed diffusion bridge to avoid drift singularities, integrating dual modulation: noise scheduling aligns diverse degradations into a shared high-entropy latent space, while path scheduling models viscous entropy regularization dynamics. The approach combines variational inference with entropy-regularized optimal transport, enabling robust single-step inference. The model explicitly decouples deterministic paths and stochastic noise, utilizing pixel-level uncertainty estimates to adaptively regulate transport trajectories, ensuring stability and generalization across heterogeneous degradations.
Key Results
- UDBM-L achieves state-of-the-art performance with an average PSNR increase of 1.5dB over existing methods across five benchmark datasets, while reducing computational complexity by 50%. It attains a 6-fold speedup in inference compared to HOGformer and DiffUIR, with fewer parameters and FLOPs. In multi-task scenarios, it demonstrates exceptional generalization, handling diverse degradations with a single model. Ablation studies confirm the effectiveness of relaxed diffusion bridges and dynamic scheduling in mitigating drift singularities and enhancing adaptation.
- On real-world benchmarks, UDBM surpasses prior models in both fidelity and perceptual quality, maintaining stable performance on unseen degradation types. Its ability to unify heterogeneous degradation manifolds into a shared high-entropy latent space enables seamless multi-task restoration, outperforming fixed-schedule and control-based approaches.
Significance
This work advances multi-task image restoration by integrating pixel-wise uncertainty into a theoretically grounded diffusion framework. It addresses longstanding issues of drift instability and rigid scheduling, providing a unified, efficient, and robust solution. The approach has broad implications for practical applications such as autonomous driving, medical imaging, and surveillance, where diverse and unpredictable degradations are common. The theoretical guarantees and single-step inference capability mark significant progress toward real-time, high-fidelity image restoration in complex environments.
Technical Contribution
The paper introduces a relaxed diffusion bridge that regularizes drift singularities by modeling degradation stochasticity via Gaussian distributions centered at degraded observations. It combines entropy-regularized optimal transport with a dual modulation strategy—noise schedule for manifold alignment and path schedule for viscous dynamics—enabling adaptive, pixel-wise control. The model guarantees Lipschitz continuity of the transport dynamics, facilitating stable, single-step inference. These innovations collectively push the boundaries of diffusion-based multi-task image restoration, offering a new theoretical and engineering paradigm.
Novelty
This is the first work to incorporate pixel-wise uncertainty into diffusion bridges for all-in-one image restoration, employing a relaxed terminal constraint to resolve drift singularities. The integration of entropy-regularized optimal transport with dynamic scheduling strategies is novel, enabling flexible, adaptive transport paths that unify heterogeneous degradation manifolds within a high-entropy latent space. These innovations significantly differ from prior fixed-schedule or coarse control methods, establishing a new framework for multi-task, high-fidelity image restoration.
Limitations
- The accuracy of pixel-wise uncertainty estimation critically affects restoration quality; errors in uncertainty prediction can lead to suboptimal transport paths, especially in extreme degradation scenarios.
- While single-step inference is efficient, its performance may degrade under highly complex or severe degradations, requiring further robustness improvements.
- Training involves sensitive hyperparameters, and generalization to unseen or highly diverse degradations still requires validation. Computational costs, though reduced, remain non-trivial for large-scale applications.
Future Work
Future research will focus on enhancing uncertainty estimation robustness, integrating multi-scale and multi-modal information, and reducing computational overhead. Exploring self-supervised and reinforcement learning strategies could further improve adaptability to unseen degradations. Extending the framework to video and 3D data, as well as real-time deployment, are promising directions to broaden practical impact.
AI Executive Summary
Image restoration remains a fundamental challenge in computer vision, especially when dealing with diverse and complex degradations encountered in real-world scenarios. Traditional methods often target specific degradation types, such as denoising or deblurring, and struggle to generalize across multiple tasks. Recent advances in diffusion models have demonstrated remarkable success in generative tasks, but their application to multi-task image restoration has been limited by rigid scheduling strategies and instability issues like drift singularities.
This paper introduces the Uncertainty-Aware Diffusion Bridge Model (UDBM), a novel framework that unifies heterogeneous degradations through pixel-wise uncertainty-guided stochastic transport. The core idea is to reformulate AiOIR as a diffusion process with a relaxed terminal constraint, which models the inherent stochasticity of degradations and avoids the drift singularities that plague standard diffusion bridges. By explicitly incorporating pixel-level uncertainty estimates, UDBM dynamically modulates the transport trajectory and noise distribution, effectively aligning diverse degradation manifolds into a shared high-entropy latent space.
The methodology leverages entropy-regularized optimal transport principles, combining a dual modulation strategy: the noise schedule ensures diverse degradations are mapped into a common high-entropy space, while the path schedule simulates viscous entropy regularization dynamics, allowing adaptive, geometry-aware refinement. This approach guarantees Lipschitz continuity of the transport dynamics, enabling high-fidelity, single-step inference across multiple tasks.
Extensive experiments on benchmark datasets such as Rain100H, Snow100K, and Haze-4K demonstrate that UDBM outperforms existing state-of-the-art methods by approximately 1.5dB PSNR, with a sixfold increase in inference speed and reduced computational cost. The model exhibits excellent generalization to unseen degradations, maintaining stable performance in real-world scenarios. Ablation studies confirm the effectiveness of the relaxed diffusion bridge and dynamic scheduling strategies in mitigating drift issues and enhancing multi-task adaptability.
Overall, UDBM represents a significant step forward in multi-task image restoration, combining strong theoretical guarantees with practical efficiency. Its ability to handle complex, heterogeneous degradations in a unified, single-step manner opens new avenues for real-time, high-quality image enhancement applications across industry and research domains. Future work will focus on further robustness, multi-scale integration, and deployment in resource-constrained environments.
Deep Analysis
Background
Image restoration技术经历了从传统滤波、变换到深度学习的快速发展。早期方法如BM3D、DnCNN主要针对单一退化类型，效果有限。近年来，扩散模型（如DDPM、Score-based）在图像生成和修复中表现出巨大潜力，但多任务场景中的应用仍受调度刚性和漂移奇异性限制。多任务学习旨在同时处理多种退化，但面临模型泛化和调度优化的难题。现有方法多依赖粗粒度控制或固定调度，难以适应复杂、多变的退化环境。
Core Problem
多任务图像恢复面临退化多样性和模型泛化的双重挑战。不同退化类型的目标冲突导致优化困难，调度策略缺乏灵活性，漂移奇异性引发训练不稳定，限制模型在复杂场景中的应用。如何设计一种既能应对多样退化，又保证推理高效稳定的模型，成为亟待解决的核心问题。
Innovation
主要创新包括：1）引入松弛扩散桥，避免漂移奇异性，增强模型稳定性；2）结合像素级不确定性，动态调节传输路径，实现多样退化信息的高效融合；3）采用双调制策略：噪声调度将退化映射到高熵潜在空间，路径调度模拟粘性动力学，优化传输轨迹；4）实现单步推理，保证模型高效性与稳定性。这些创新共同推动多任务图像恢复迈向更高性能与鲁棒性。
Methodology
- �� 将AiOIR转化为像素级不确定性引导的随机传输问题。
- �� 引入松弛扩散桥，使用高斯分布替代严格终端约束，缓解漂移奇异性。
- �� 结合Entropy-Regularized Optimal Transport，设计噪声与路径调度。
- �� 噪声调度将不同退化映射到高熵潜在空间，路径调度模拟粘性动力学，动态调节传输轨迹。
- �� 采用像素级不确定性估计，调节调度参数，实现多任务适应。
- �� 利用变分推断与线性SDE，简化训练与推理流程，保证单步高效。
Experiments
在五个公开数据集（如Rain100H、Snow100K、Haze-4K）上，训练模型并与SOTA方法（如HOGformer、DiffUIR）对比。指标包括PSNR、SSIM、感知质量指标（如MANIQA）。采用不同退化强度（噪声、模糊、雾霾）进行多任务训练，验证模型泛化能力。还进行了消融实验，分析松弛扩散桥和调度策略的贡献。模型参数、计算量、推理速度均优于对比方法。
Results
UDBM-L在五个任务中平均PSNR提升1.5dB，SSIM提升0.02，推理速度比HOGformer快6倍，参数减少50%。在未见退化类型上表现出优异的泛化能力，单模型适应多种场景。消融实验显示，松弛扩散桥显著缓解漂移奇异性，动态调度增强复杂退化的适应性。多任务场景中，模型保持稳定性能，优于传统调度策略。
Applications
模型适用于自动驾驶、医疗影像、监控监测等多场景图像修复任务，能在单次推理中完成多种退化的高质量恢复。其高效性使其适合实时处理需求，减少硬件成本。未来可结合多模态信息，提升复杂环境下的鲁棒性，推动行业智能化升级。
Limitations & Outlook
模型对不确定性估计的准确性敏感，极端退化场景下仍存在性能瓶颈。单步推理虽快，但在极复杂退化环境中效果有限。训练过程中超参数调节复杂，泛化能力在未训练场景中仍需验证。未来需优化不确定性估计机制，降低计算成本，增强鲁棒性。
Plain Language Accessible to non-experts
想象你在厨房做饭，面对各种食材和调料，有时食材新鲜，有时变质，有时调料用多了，有时用少了。传统厨师只会专注一种食材或调料，做出来的菜只能应付特定场景。现在，假设你有一台神奇的厨师机器人，它可以根据每次食材状态自动调整烹饪策略，既能处理新鲜食材，也能应对变质或调料不足的问题。它通过观察每个食材的状态，动态调整火候和调料比例，确保菜肴美味。这台机器人就像UDBM一样，能在不同图像退化情况下，自动调节修复策略，快速高效还原清晰图像。它用一种智能“感知”机制，判断每个部分的退化程度，灵活调节修复路径，避免传统方法中的“卡壳”或“崩溃”。这样，无论图像多复杂、多变，这个系统都能像个聪明的厨师一样，快速做出令人满意的“菜肴”。
ELI14 Explained like you're 14
想象你在玩一款游戏，你的角色有时候会遇到各种问题，比如迷路、掉线、被攻击或迷失方向。以前的游戏只能解决一种问题，比如帮你找到路，但不能同时应对掉线或被攻击的问题。现在，有一种超级智能的游戏助手，它可以根据每次遇到的问题，自动调整策略，帮你快速解决所有困难。这个助手会观察你的情况，判断你遇到的具体问题，然后用最快的方法帮你恢复正常。它就像一位聪明的朋友，能在你遇到各种难题时，立刻给出最合适的建议。它用一种特别的“感知”能力，知道哪里出了问题，怎么修复，甚至提前预警。它的核心技术就是根据每个问题的严重程度，动态调整自己的行动路线，确保你能在最短时间内顺利完成任务。就像你有个万能的朋友，总是在你最需要的时候出现，帮你解决所有难题。这就是UDBM的核心思想：用智能感知和动态调整，让复杂的问题变得简单快速。
Abstract
All-in-One Image Restoration (AiOIR) faces the fundamental challenge in reconciling conflicting optimization objectives across heterogeneous degradations. Existing methods are often constrained by coarse-grained control mechanisms or fixed mapping schedules, yielding suboptimal adaptation. To address this, we propose an Uncertainty-Aware Diffusion Bridge Model (UDBM), which innovatively reformulates AiOIR as a stochastic transport problem steered by pixel-wise uncertainty. By introducing a relaxed diffusion bridge formulation which replaces the strict terminal constraint with a relaxed constraint, we model the uncertainty of degradations while theoretically resolving the drift singularity inherent in standard diffusion bridges. Furthermore, we devise a dual modulation strategy: the noise schedule aligns diverse degradations into a shared high-entropy latent space, while the path schedule adaptively regulates the transport trajectory motivated by the viscous dynamics of entropy regularization. By effectively rectifying the transport geometry and dynamics, UDBM achieves state-of-the-art performance across diverse restoration tasks within a single inference step.