Resfusion: Denoising Diffusion Probabilistic Models for Image Restoration Based on Prior Residual Noise

Recently, research on denoising diffusion models has expanded its application to the field of image restoration. Traditional diffusion-based image restoration methods utilize degraded images as conditional input to effectively guide the reverse generation process, without modifying the original denoising diffusion process. However, since the degraded images already include low-frequency information, starting from Gaussian white noise will result in increased sampling steps. We propose Resfusion, a general framework that incorporates the residual term into the diffusion forward process, starting the reverse process directly from the noisy degraded images. The form of our inference process is consistent with the DDPM. We introduced a weighted residual noise, named resnoise, as the prediction target and explicitly provide the quantitative relationship between the residual term and the noise term in resnoise. By leveraging a smooth equivalence transformation, Resfusion determine the optimal acceleration step and maintains the integrity of existing noise schedules, unifying the training and inference processes. The experimental results demonstrate that Resfusion exhibits competitive performance on ISTD dataset, LOL dataset and Raindrop dataset with only five sampling steps. Furthermore, Resfusion can be easily applied to image generation and emerges with strong versatility. Our code and model are available at https://github.com/nkicsl/Resfusion.

Paper

References (65)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer KnJY6/10 · confidence 4/52024-07-05

Summary

The paper presents Resfusion that leverages prior residual noise to improve restoration performance. It introduces a smooth equivalence transformation for learning residual noise and demonstrates the efficacy of Resfusion through extensive experiments and ablation studies.

Strengths

The paper is well-written and the idea is clearly illustrated. The technique of finding T' to simplify Eq 6 is clever. The experiments are detailed and show non-trivial improvements.

Weaknesses

1. Since T' is smaller than the usually used T=1000, the sampling process of the proposed method seems to be faster than previous methods. But there are no comparisons on inference speed. 2. What is the scheduling of \alpahs? The authors are suggested to add more experimental analysis about the error introduced by ignoring terms in Eq 6.

Questions

See the weakness.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors have discussed the limitations clearly.

Reviewer 4Jy95/10 · confidence 4/52024-07-12

Summary

This paper presents a general diffusion framework for image restoration, named Resfusion. The main idea is to introduce a residual term to DDPM to directly generate clean images from degraded images. Moreover, the form of the inference process is consistent with the DDPM and allows very few sampling steps. The model is evaluated on ISTD, LOL, and Raindrop dataset. Several ablation experiments on residual terms and loss functions are also conducted. In discussion, the authors show that their method can be used for general image generation by set $\hat{x}_0$ to 0.

Strengths

- The idea of adding a residual term to the diffusion process is interesting. - The proposed smooth equivalence transformation is promising. - The experimental results are good.

Weaknesses

- How do you get/design Eq. (3)? I wonder if the term $(1-\sqrt{\alpha_t})R$ is manually designed or derived from a specific equation. - In Eq. (6), $\beta_t$ is not defined in the context. - I can understand that Eq. (8) aims to ensure the coefficient of $x_0$ in Eq. (7) is close to 0. But how can you obtain Eq. (9) in which $\hat{x}_0$ is non-zero? - The overall derivation in Section 2.2 is unclear and confusing. Please make the connections between Equations more smooth. In Eq. (12), how do you get the variance $\Sigma_\theta$? And how do you obtain Eq. (13)? Please explain why and add the reference papers. - In experiments, it would be better to unify the evaluation metrics. For example, using PSNR, SSIM, LPIPS for all tasks and datasets (FID or MAE can be also added). - The presentation (especially the derivation) should be improved.

Questions

The proposed method seems to be sensitive to noise. Can you train the model for the denoising tasks?

Rating

5

Confidence

4

Soundness

3

Presentation

2

Contribution

3

Limitations

Please see the Weaknesses.

Reviewer 4Jy92024-08-09

I appreciate the authors' efforts in the rebuttal. For the notations and definitions in the paper, I want to note that although you use the same notations from other papers (e.g., DDPM), you still need to define/clarify them again in your draft so that the readers can understand them correctly (otherwise it would be confusing). I believe the main idea of this paper is similar to ResShift and RDDM. However, in the rebuttal, the derivation in Section 2.2 is still unclear and there is no answer to how to obtain Eq. (9). Therefore, I choose to maintain my original score i.e. * Borderline accept*.

Authorsrebuttal2024-08-09

Thank you very much for the patient response and constructive suggestions. Below we address specific questions and comments. > the derivation in Section 2.2 is still unclear We provide a detailed logical relationship for the equations in Section 2.2 in our rebuttal: - To obtain a computable starting point, we introduce the smooth equivalence transformation technique, corresponding to Eq. (7) - Eq. (9). - Because both the forward and backward processes only involve $t \le T'$, we unify the training and inference processes, corresponding to Eq. (10) - Eq. (14). - We explained the working principle of Resfusion from the perspective of vector intersection, corresponding to Eq. (15). > there is no answer to how to obtain Eq. (9) Since $\sqrt{\overline{\alpha}_{T'}}$ is closed to 0.5 as in Eq. (8), $2\sqrt{\overline{\alpha}_{T'}}-1$ is closed to zero in Eq (7). Then we can derive Eq. (9) from Eq. (7), where $\hat{x}_{0}$ represents the input degraded image.

Area Chair Ydph2024-08-10

Thank you for the discussion. Any other thoughts from other reviewers?

Reviewer 4Jy92024-08-13

Thanks. The second question is clear now. Please add these explanations (in all rebuttals) to the revised draft which would definitely improve the readability of your work. From my side, the overall paper is on the borderline so I will keep my original score.

Authorsrebuttal2024-08-13

Thank you very much for the valuable feedback. We greatly appreciate your recognition of our work. We will certainly provide more detailed explanations in the revised paper.

Reviewer DA7x5/10 · confidence 3/52024-07-13

Summary

This paper proposes to start the reverse diffusion process from the noisy degraded images for image restoration. It introduces a weighted residual noise as the prediction target and leverage a smooth equivalence transformation to find the starting noise. The experiments shows competitive performance on shalow removal, low-light enhancement and deraining, with shortened sampling steps.

Strengths

1, It predicts the residual noise to allow for diffusion directly from the noisy degraded images 2, It transforms the learning of the noise term into the renoise term and follows the same inference process as DDPM 3, Shortened inference steps without redesigning the noise schedule; sota results

Weaknesses

1, ResShift also shifts toward the residual term. What are the advantages and disadvantages compared with ResShift? 2, Visualization of the five inference steps is lacking. 3, What is the results with the complete inference steps (e.g., T=1000)? 4, In discussion, what is the model size? Can a 7.7M model achieve such good results in Figure 7?

Questions

See weakness.

Rating

5

Confidence

3

Soundness

2

Presentation

2

Contribution

2

Limitations

See weakness.

Authorsrebuttal2024-08-13

Thank you very much for the recognition of our work and rebuttal. We are delighted to have addressed your concerns. We would greatly appreciate it if you could consider raising the score.

Reviewer vAm86/10 · confidence 3/52024-07-18

Summary

This paper proposes a method that leverages generative diffusion for image restoration tasks. The authors suggest incorporating the residual term, defined as the difference between the corrupted and clean images, into the forward and reverse diffusion processes. The forward process is described by a Markov chain, where the probability of the image at each time step $t$ is conditioned on the previous step, $t - 1$, and the residual term. This results in a forward where at each time step $t$, the image obtained at the previous time step $t - 1$ is mixed with the residual term and an additional white Gaussian noise. Consequently, the image at each time step can be described as a weighted sum of three components: the clean image, the residual term, and an additive white Gaussian noise. Since the residual term is the difference between the corrupted and clean images, this weighted sum can be rewritten as a weighted sum of the clean image, corrupted image, and Gaussian noise. The forward process stops when the weight of the clean image becomes approximately zero. Thus, the forward process gradually transfers a clean image into a weighted sum of the corrupted image and white Gaussian noise. Correspondingly, the reverse process starts from this weighted sum and incrementally reduces the corruption and noise. Because the corrupted image is available, the initialization for the reverse process is easily obtained by mixing the corrupted image with white Gaussian noise. The authors derive the expression for the mean of the reverse process and train a network to predict, at each step, the difference between the corrupted noisy image obtained at that step and the clean image. The proposed algorithm evaluated thorough experiments on three reconstruction tasks: shadow removal, low light enhancement, and deraining, comparing the performance with recent competing methods. In all experiments, the authors apply five diffusion steps. The results show that the proposed algorithm generally outperforms the competitors while consuming fewer computational resources (multiplication operations times the number of diffusion steps).

Strengths

I believe the idea of conditioning on the residual term is original and significant. It allows deriving a reverse process that starts from the corrupted image mixed with noise rather than from pure noise, reducing the number of diffusion steps required for image restoration.

Weaknesses

- Several notations are used without being defined. For example, the definitions of $\alpha_t$, $\beta_t$, $\bar{\alpha}_t$ are missing. Although the paper generally follows the notations defined in [1], defining these notations in the current paper would make it easier for readers to follow the explanations. - Some mathematical derivations are not detailed enough. For example, the proof in Appendix A.1 is not sufficiently detailed. Another example is derivation of $\mu_{\theta}$ in Equation (13), which is missing. It is completely fine to rely on derivations in [1] or [2], but I believe including the relevant parts of the derivations in the appendix would greatly help readers follow the derivations. [1] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840–6851, 2020. [2] Calvin Luo. Understanding diffusion models: A unified perspective. arXiv preprint arXiv:2208.11970, 2022.

Questions

To provide a more comprehensive background, it might be beneficial to cite [3] which introduced a general framework for leveraging generative diffusion for solving image restoration problems where the observed image is contaminated by linear degradation and additive white Gaussian noise. [3] Bahjat Kawar, Gregory Vaksman, and Michael Elad. SNIPS: Solving noisy inverse problems stochastically. Advances in neural information processing Systems 34:21757-21769, 2021.

Rating

6

Confidence

3

Soundness

3

Presentation

2

Contribution

3

Limitations

Yes.

Area Chair Ydph2024-08-09

Dear reviewers, do the authors' responses answer your questions or address your concerns? Thanks.

Area Chair Ydph2024-08-12

Dear reviewers, as we approach the final two days, please take a moment to review the author's responses and join the discussion. Thank you!

Area Chair Ydph2024-08-14

Dear reviewers, the authors are eagerly awaiting your response. The author-reviewer discussion closes on Aug 13 at 11:59 pm AoE. Thanks!

Authorsrebuttal2024-08-14

We sincerely appreciate Area Chair Ydph. Your responsible feedback has greatly inspired us. We would also like to express our gratitude to all the reviewers for their professional review and recognition of our work.

Reviewer vAm82024-08-13

Thank you for the answers. After carefully considering the comments from other reviewers, the authors' feedback, and the discussion with reviewer 4Jy9, I have decided to maintain my original rating, i.e., Weak Accept.

Authorsrebuttal2024-08-13

Thank you sincerely for the professional review and recognition of our work.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC