IRBridge: Solving Image Restoration Bridge with Pre-trained Generative Diffusion Models

We are grateful for your positive review! We will carefully address your concerns below.

Q1: Empirical selection of hyperparameters

We acknowledge that this is a limitation of IRBridge, and we offer two key insights.

• As shown in Appendix C, while different choices of timesteps can significantly affect the intermediate states of the diffusion trajectory, their impact on final results is generally limited. This suggests that although we may not adopt theoretically optimal inference hyperparameters, the actual performance loss is often acceptable (~5%). Based on our experience, the hyperparameters used for the inpainting task are effective for most tasks, and we refer it as a default setting.
• Recent works have provided ways to systematically optimize hyperparameters under certain metrics. For example, LDSB [1] recently proposed using local Schrödinger bridge to optimize the diffusion coefficients. IRBridge can benefit from it to obtain optimal hyperparameters in the context of relative entropy minimization.

While optimal hyperparameters are attainable, we believe that the cost of complex optimization is not cost-effective, as manually designed parameters are typically sufficient.

Q2: Inference efficiency and computational resources.

In our implementation, we use the SD1.5 model, which processes approximately 7.14 steps per second on a 512×512 image using an RTX 3090 GPU, consuming around 6300MB of VRAM. In THIS TABLE , we present the performance and inference time of IRBridge under different inference steps. Encouragingly, even with a 75% reduction in inference steps, IRBridge shows minimal performance drop (less than 3%) while significantly reducing inference time (14.2s → 3.4s), demonstrating its potential for improving inference efficiency. Future work can also focus on compression of the pretrained model to further enhance the inference efficiency.

Q3: Generalization for real-word degradation.

We present the evaluation results on real-world datasets in THIS TABLE. We utilized no-reference image quality metrics to evaluate the performance of the methods. As shown in the table, IRBridge demonstrates a clear advantage over other training-from-scratch methods in real-world scenarios. Compared to the latest method, DCPT (ICLR 2025), IRBridge still achieves a better performance (with average 5.6% performance improved), highlighting its superior performance under real-world conditions. We attribute these improvements to the pretrained generative prior that equips the model with stronger generalization capability. Visual examples are provided HERE.

Q4: Has the work considers reducing the dependence on the initial state estimation by introducing additional prior information from different perspectives?

We would like to clarify that, mathematically the dependence on the initial state is solely determined by β in Eq.7. In principle, we only consider the case where β takes its minimum value, as this theoretically implies minimal reliance on x0.

However, we are open to the idea of retaining some prior information to reduce dependence on x0. For example, given a known degradation model, a plug-and-play (PnP) [2] style approach can be adopted to enforce observation consistency at each iteration step, thereby reducing the degrees of freedom in estimation. Similarly, in the case of image inpainting, we can follow the strategy of RePaint [3] by directly replacing the corresponding pixels in the generated x0 with the uncorrupted pixels from the original LQ sample, thus lowering the reliance on the model’s estimate of initial state.

Q5: Can the estimation errors of the initial state be gradually corrected through iterative optimization?

Diffusion-based image generation and restoration methods can generally be viewed as progressively refining their estimate of the initial state during the reverse iteration process. IRBridge, by nature, leverages a pretrained generative model to simulate the reverse process of a restoration bridge. As such, its inference procedure inherently serves to iteratively correct the estimation error of the initial state.

Q6: Inconsistences in expression and the quality of images and tables.

We sincerely appreciate the reviewer’s thoughtful suggestions! We will address these issues in the revised version. These constructive suggestions are invaluable for enhancing the overall quality of our work.

[1] Finding Local Diffusion Schrodinger Bridge using Kolmogorov-Arnold Network, CVPR 2025

[2] Denoising Diffusion Models for Plug-and-Play Image Restoration, CVPR2023

[3] RePaint: Inpainting using Denoising Diffusion Probabilistic Models, CVPR 2022

We sincerely appreciate your insightful review and valuable comments! We will carefully address your concerns below.

Q1: Insufficient comparative experiments.

We have added a QUANTITATIVE COMPARISON between IRBridge and the suggested methods across different tasks. Since the training code for DCPT has not been released, we directly used the pretrained model provided by the authors for comparison. All results are reported at a unified resolution of 512×512 to ensure a fair comparison.

The proposed IRBridge outperforms existing methods on most tasks. While it may underperform on deterministic metrics like PSNR and SSIM due to its generative nature and output diversity (e.g., slight tonal shifts from ground truth), these variations rarely affect perceptual quality, though they impact sensitive metrics. In contrast, IRBridge consistently achieves better FID scores (lower average ~4.97), highlighting its strength in capturing data distribution. Moreover, we present the quantitative results of IRBridge on real-world datasets in THIS TABLE. IRBridge outperforms the latest method DCPT, achieving an average performance improvement of approximately 5.6%. These results demonstrate the generalization ability of IRBridge.

Q2: The problem tackled in this paper appears somewhat gradual.

We thank the reviewer for their thoughtful comments, and we provide insights from both theoretical and practical perspectives.

• Theoretically. The mentioned methods, similar to IRSDE, can be regarded as defining a stochastic differential equation (SDE) that describes the transformation from LQ images to HQ counterparts. They typically train models to estimate the score of the corresponding reverse SDE for image restoration. In contrast, our method explores the use of score estimators from generative SDEs (pretrained DMs) to estimate the score of these SDEs. By leveraging a generative model capable of estimating $p(x_{0} | x_t^*, x_{lq})$, IRBridge can directly solve the reverse SDE defined by methods like IRSDE and GOUB.
• Practically. Our method is the first to enable bridge methods to directly leverage pretrained DMs, whose priors not only reduce the training burden but also provide better generalization (+5.6% vs DCPT in real-world scenarios) , as demonstrated in Q1. Moreover, IRBridge effectively decouples training and inference: by leveraging a DM conditioned on LQ images, IRBridge can solve alternative forms of restorative SDEs, without being constrained to specific diffusion coefficients, enabling greater flexibility.

Q3: The role of ControlNet and training issues.

The employed ControlNet integrates guidance information from the degraded image, ensuring that the model is directed to generate the desired x0. From another perspective, ControlNet can indeed be seen as performing implicit image restoration, as it models $p(x_{hq} | x_t^*, x_{lq})$. However, we emphasize a fundamental distinction: ControlNet is trained via conditional score matching, following standard diffusion. Our ControlNet is trained to estimate the score of the SDE associated with a generative diffusion, whereas restoration bridge methods like GOUB are trained to estimate the score of their restorative diffusion . Moreover, these methods also use LQ images as condition for the model to enhance performance.

We clarify that IRBridge does not focus on the training of the model itself, yet we acknowledge that its performance depends on the underlying model. The ControlNet we used is a relatively coarse implementation, and other conditional control methods (like UniCon[1]) remain applicable to IRBridge, for fine-tuning the generative model. Future work can explore more effective conditional guidance strategies and guidance modules tailored for IR tasks.

Q4: Training an all-in-one image restoration bridge cannot be considered a highlight contribution.

We agree that training an all-in-one model on multiple degradation datasets can achieve good results. However, we want to clarify that the main contribution of IRBridge lies in enabling the solution of IR bridges through pre-trained DMs for the first time, rather than training an all-in-one model. Typically, training an all-in-one model requires merging multiple datasets to handle diverse image types, leading to high computational costs. In contrast, IRBridge allows direct leverage of generative priors, significantly lowering the training effort (discussed in Section 4.2).

Q5: Table quality and missing details.

The desnowing dataset contains 50k samples in both the training and testing sets. We will correct the visual quality of the tables and figures and address the missing details in the revised version. Thanks pointing that out!

[1] A Simple Approach to Unifying Diffusion-based Conditional Generation, ICLR 2025

We sincerely appreciate your insightful comments! We will address your concerns point by point.

Q1: Evaluations on real-world scenarios.

We present the EVALUATION RESULTS on real-world datasets (RealRain-1K for deraining, RealSnow for desnowing, and RTTS for dehazing). As these datasets lack ground truth, we utilized widely-used no-reference image quality metrics (MUSIQ, BRISQUE, NIQE) to evaluate the performance of the methods. As shown in the table, IRBridge demonstrates a clear advantage over other training-from-scratch methods in real-world scenarios. Compared to the latest method, DCPT (ICLR 2025), IRBridge still achieves a better performance (with average 5.6% performance improved), highlighting its superior performance under real-world conditions. We attribute these improvements to the pretrained generative prior that equips the model with stronger generalization capability. Visual examples are provided HERE.

Q2: Inference efficiency.

We acknowledge that the iterative nature of IRBridge makes it less suitable for real-time processing. However, its high flexibility allows for improving efficiency by reducing the number of inference steps. We present the performance and inference time of IRBridge under different numbers of inference steps (on a 3090 GPU processing standard 512×512 resolution images). The TABLE shows that even with a 75% reduction in inference steps, the performance degradation remains minimal (< 3%), demonstrating it’s potential for lowering inference overhead. Additionally, future work could focus on distillation and compression of the pretrained generative model to further reduce the computational complexity of IRBridge and improve its inference efficiency.

Q3: Manual Hyperparameter Tuning.

We acknowledge that it is a limitation of IRBridge. To address this, we offer two insights:

• 1) Manual configuration as a practical solution. In Appendix C, we show the impact of timestep selection strategies across different tasks. Our experiments indicate that although timestep selection impacts intermediate states, it has a minimal effect (~5%) on final performance, suggesting manual tuning is both sufficient in most scenarios. Based on our experience, the timestep schedule used in inpainting task tends to work well for most tasks, which is referred as the default setting.

• 2) Systematic hyperparameter selection. Several approaches can provide more principled strategies for selecting optimal parameters under specific considerations. For example, the recent work LDSB [1] proposes using a path prediction network to estimate more suitable diffusion coefficients via a local Schrödinger bridge, thereby improving the performance of pretrained DMs. Similarly, for IRBridge, one could explore optimal timestep scheduling under the context of minimizing relative entropy.

For a trade-off between cost and performance, we still recommend using manually selected parameters. Pursuing complex optimization for marginal performance gains is not cost-effective.

Q4: How does the transition equation differentiate IRBridge from other approaches, and in what ways does it offer an advantage over diffusion-based IR methods?

We would like to elaborate on the contributions of IRBridge from both theoretical and practical perspectives.

• 1) Theoretically. IRBridge supports bridging two distinct diffusion processes that share the same endpoints. In contrast to previous IR bridge methods, IRBridge does not define a specific diffusion process tailored for restoration. Instead, it leverages the transition equation to ”substitute“ a pretrained DDPM model in place of a specially trained score network for solving the corresponding SDE. To our knowledge, this is the first work to introduce a pretrained generative model into such bridge-based restoration frameworks, breaking the limitation of prior approaches that require training models from scratch.
• 2) Practically. The incorporation of pretrained generative priors endows IRBridge with stronger generalization capabilities, especially in complex scenes or across different data domains. In Q1, we demonstrated the superior generalization ability of IRBridge compared to other bridging methods in real-world scenarios. Additionally, as discussed in Section 4.2, the pretrained model provides a strong initialization, enabling IRBridge to achieve nearly twice the training efficiency compared to methods trained from scratch (e.g., GOUB).

Q5: Correction of some typos.

We will thoroughly check for spelling errors in the subsequent revision to improve the quality of our manuscript. Your feedback is invaluable for improving our work!

[1] Finding Local Diffusion Schrodinger Bridge using Kolmogorov-Arnold Network, CVPR 2025

We are grateful for your valuable comments, which are helpful for improving our manuscripts. Below, we will address your concerns point by point.

Q1: Test image resolution and supported resolution.

To ensure a fair comparison, we conducted all comparison experiments at a 512×512 resolution in our paper. However, it should note that the resolution supported by IRBridge depends on the model it employs. For SD1.5’s UNet, it supports resolutions in multiples of 64. Inspired by WeatherDiff, we developed an overlapping patches partitioning scheme to support inputs of arbitrary resolutions. We present VISUAL EXAMPLES of IRBridge processing arbitrarily high-resolution real-world images, highlighting its ability to effectively handle high-resolution images without causing performance degradation.

Q2: Inference efficiency of IRBridge.

With the default setting of 100 inference steps, IRBridge takes about 14.2 seconds to process a 512 resolution image, running at 7.18 steps per second. Notably, since IRBridge decouples model training from inference, it allows flexible specification of both the diffusion coefficients and the number of inference steps. In THIS TABLE , we present the inference time and quantitative results of IRBridge under different step settings. It is worth noting that even with a 75% reduction in inference steps (14.2s -> 3.4s of inference time), IRBridge still maintains comparable performance (<3% average drop), demonstrating its potential for more efficient inference.

Q3: The performance gains from pretrained diffusion priors.

The performance gains brought by pretrained priors can be attributed to three main aspects:

• 1) Improved Generalization. Pretrained models provide general knowledge about a wide range of image distributions, enhancing the model’s ability to generalize to various scenarios. IRBridge demonstrates better generalization across different data domains compared to methods trained from scratch (as shown in Fig.7). Furthermore, we present the results of IBridge on real-world datasets (see Q4), showing its superior performance compared to other methods, further supporting our claims.
• 2) Faster Convergence. As shown in Section 4.2, IRBridge benefits from a favorable initialization provided by the pretrained model, leading to nearly twice the training efficiency compared to training-from-scratch methods such as GOUB.
• 3) Better Robustness. Even though we simply adopted a ControlNet trained with a DDPM objective, its performance still outperforms methods trained specifically to estimate the score of a particular SDE. This demonstrates that the pretrained generative prior provides IRBridge with more robust representations, thereby enhancing its overall performance.

Q4: Comparison on more datasets and with the latest methods.

We provide QUANTITATIVE RESULTS of IRBridge compared with recent methods, including ResShift [1], RDDM [2], DiffUIR [3], and DCPT [4], across different tasks. Additionally, we present the QUANTITATIVE RESULTS of IRBridge and the aforementioned methods on real-world degraded datasets, covering image deraining, desnowing, and dehazing.

IRBridge outperforms other bridging methods, achieving an average improvement of approximately 3.23% over the latest bridge method DiffUIR, demonstrating that the pretrained generative prior significantly enhances model robustness. Furthermore, IRBridge achieves an overall improvement of about 5.60% over the recent method DCPT on real-world scenarios, confirming that the introduced prior in IRBridge improves the model's generalization across different degradation conditions. We present VISUAL EXAMPLES of IRBridge processing real-world degraded images.

Q5: The size of the test data is not specified.

We present the complete number of training and testing samples for the datasets used in the table below.

We thank the reviewer for the thorough review. These detailed suggestions are invaluable for improving the quality of our manuscript.

[1] ResShift: Efficient Diffusion Model for Image Super-resolution by Residual Shifting. NeurIPS 2023

[2] Residual Denoising Diffusion Models. CVPR 2024

[3] Selective Hourglass Mapping for Universal Image Restoration Based on Diffusion Model. CVPR 2024

[4] Universal Image Restoration Pre-training via Degradation Classification. ICLR 2025.