acceptodds
Under review as a conference paper at ICLR 2027

SDDBMs: Soft Denoising Diffusion Bridge Models

Abstract

Diffusion bridge models leverage Doob's \(h\)-transform to construct stochastic transports between arbitrary endpoint distributions, and have shown strong potential in image-to-image translation and restoration. However, most existing bridge models rely on hard endpoint conditioning, which forces the terminal state to match a prescribed target exactly. This hard constraint induces terminal-boundary singularities: the terminal law collapses to a Dirac measure, and the resulting drift coefficients become ill-conditioned near the endpoint. In this paper, we propose Soft Denoising Diffusion Bridge Models (SDDBMs), a generalized framework that regularizes diffusion bridges directly at the level of their terminal constraints. Instead of imposing an exact endpoint, SDDBMs prescribe a non-degenerate Gaussian terminal marginal under the transformed path measure, with a flexible terminal center and variance. Starting from this prescribed marginal, we develop a complete closed-form construction of the soft bridge, including the Gaussian terminal reweighting and soft \(h\)-function, the induced Gaussian forward marginals and \(x_0\)-free dynamics. Theoretically, SDDBMs provide a unified probabilistic perspective that encompasses existing diffusion bridge models, including DDBMs, GOUB, and UniDB, as special cases under specific parameter choices. Extensive experiments on image restoration tasks demonstrate that SDDBMs achieve improved numerical stability and superior generation quality over existing bridge-based methods.

Primary area:
generative models
Keywords:
diffusion models, bridge models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ZtZpOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper points out and proved the numerical singularities of hard endpoint (h-transform-based) bridges from the perspective of probability theory. Based on it, the authors prescribe the non-degenerate Gaussian terminal marginal to overcome the singularity. They construct a fully closed-form, singularity-free, and x0x_0-free diffusion bridge framework and unify some previous bridges like DDBMs, GOUB, and UniDB. Three different image restoration experiments demonstrate SOTA performance of SDDBMs.

Critical strengths
  1. The paper is theoretical and the detailed proofs of the Theorems and Propositions are complete and clearly described in the Appendix. It provides an explanation of UniDB via probabilistic interpretation.

  2. SDDBMs demonstrate superior performance across different image restoration tasks under four standard metrics.

Critical weaknesses
  1. Some experimental details might not be aligned well with the proposed theory. A central contribution of SDDBMs is the proposed non-degenerate terminal distribution p(xT∣x0,x⋆)=N(aTx0+bTx⋆+cTμ,σ2I)p(x_T\mid x_0,x^\star)=N(a_Tx_0+b_Tx^\star+c_T\mu,\sigma^2I). However, their experiments were conducted by initializing the reverse process with xT≈(bT+cT)x⋆x_T\approx(b_T+c_T)x^\star, which omits the unavailable source-dependent term x0x_0 and the terminal noise (Appendix H). Dropping unavailable x0x_0 does make sense since it is unavailable, but the lack of justification for the terminal noise would weaken the paper's contribution.

  2. The proposed SDDBMs framework recovers DDBMs, but the authors provided no theoretical derivations on the formulation of SDDBMs on VE, VP process and no empirical evidence on whether SDDBMs could improve original DDBMs.

Questions for authors
  1. The qualitative visual comparisons with baselines (GOUB and UniDB) are somewhat limited. The results among GOUB, UniDB, and the proposed SDDBMs should be compared in Fig. 3. More qualitative visual comparison examples would better emphasize the strengths of SDDBMs. Moreover, some examples in Fig. 3 and Fig. 8 are redundant. The appendix should provide additional diverse visual comparisons.

  2. The Number of Function Evaluations (NFEs) or sampling steps for their SDDBMs and baselines are missed. Could the authors clarify the NFEs used during inference?

  3. Proposition 3.1 would benefit from a rewriting. As it is, the main idea is to show the singularity of diffusion bridges. I could understand the injection (mT(x0)≠mT(y0)m_T(x_0) \neq m_T(y_0) for any x0≠y0x_0 \neq y_0) leads to the conclusion. However, in this version of Proposition 3.1, neither singularity nor the relation to diffusion bridges are emphasized.

  4. Lack of quantitative evaluation for the ablation on hyperparameter β\beta. Although the authors have claimed that ``different β\beta values lead to comparable performance", the specific results should be also provided to validate the claim, which would be more convincing.

10 comments

W1:Some experimental details might not be aligned well with the proposed theory. A central contribution of SDDBMs is the proposed non-degenerate terminal distribution p(xT∣x0,x⋆)=N(aTx0+bTx⋆+cTμ,σ2I)p(x_T|x_0,x^\star)=\mathcal{N}(a_Tx_0+b_T x^\star+c_T \mu, \sigma^2 \mathrm{I}). However, their experiments were conducted by initializing the reverse process with xT≈(bT+cT)x⋆x_T \approx (b_T+c_T) x^\star, which omits the unavailable source-dependent term x0x_0 and the terminal noise (Appendix H). Dropping unavailable x0x_0 does make sense since it is unavailable, but the lack of justification for the terminal noise would weaken the paper's contribution.

Thank you for this insightful question. We apologize for the insufficient explanation of the terminal initialization strategy in the original manuscript. Below, we provide a detailed explanation of why omitting the terminal noise during inference is reasonable and mathematically justified.

Under our experimental setting μ=x⋆\mu=\mathbf{x}^{\star}, the terminal distribution of SDDBMs is given by N(aTx0+(bT+cT)x⋆,σ2I)\mathcal{N}\left(a_T\mathbf{x}_0+(b_T+c_T)\mathbf{x}^{\star},\sigma^2\mathrm{I}\right). During inference, x0\mathbf{x}_0 is unavailable, and its coefficient aTa_T is sufficiently small under our experimental parameter settings (∣aT∣≤0.0036|a_T|\leq 0.0036). We therefore neglect the source-dependent contribution aTx0a_T\mathbf{x}_0 and consider the approximate terminal distribution p~T(xT∣x⋆)=N((bT+cT)x⋆,σ2I).\tilde{p}_T(\mathbf{x}_T|\mathbf{x}^\star)=\mathcal{N}((b_T+c_T)\mathbf{x}^\star,\sigma^2 \mathrm{I}). In image-to-image tasks, the degraded image x⋆\mathbf{x}^{\star} serves as guidance for the reverse process, which starts from a terminal state xT\mathbf{x}_T and aims to recover the corresponding clean image x0\mathbf{x}_0. Therefore, we need to determine a terminal state xT\mathbf{x}_T that best represents the terminal distribution. Under the expected squared error criterion, the optimal choice is xT=(bT+cT)x⋆\mathbf{x}_T=(b_T+c_T)\mathbf{x}^{\star}.

Specifically, let m=(bT+cT)x⋆\mathbf{m}=(b_T+c_T)\mathbf{x}^{\star} and xT∼N(m,σ2I).\mathbf{x}_T \sim \mathcal{N}(\mathbf{m},\sigma^2 \mathrm{I}). We seek the terminal state z\mathbf{z} that minimizes the expected squared error with respect to the terminal random variable XT\mathbf{X}_T:

z⋆=arg⁡min⁡zEp~T[∣∣XT−z∣∣22]\mathbf{z}^{\star}=\arg \min_{\mathbf{z}} \mathbb{E}_{\tilde{p}_T}\left[||\mathbf{X}_T-\mathbf{z}||_2^2\right] Using the Gaussian reparameterization xT=m+σϵ\mathbf{x}_T=\mathbf{m}+\sigma\boldsymbol{\epsilon}, where ϵ∼N(0,I)\boldsymbol{\epsilon}\sim\mathcal{N}(\mathbf{0},\mathrm{I}), and noting that E[ϵ]=0\mathbb{E}[\boldsymbol{\epsilon}]=\mathbf{0} and E[∣∣ϵ∣∣22]=d\mathbb{E}\left[||\boldsymbol{\epsilon}||^2_2\right]=d, we obtain

\left[ ||\mathbf{X}_T-\mathbf{z}||_2^2 \right] =||\mathbf{z}-\mathbf{m}||_2^2+d\sigma^2,$$ where $d$ denotes the dimensionality of the image. Since $d\sigma^2$ is independent of $\mathbf{z}$, the expected squared error is uniquely minimized when $\mathbf{z}=\mathbf{m}$. Therefore, we initialize the reverse sampling process with $\mathbf{x}_T=(b_T+c_T)\mathbf{x}^\star$, which is the optimal point estimate of the terminal distribution under the expected squared error criterion. We have also provided a detailed explanation of this initialization strategy in the revised manuscript. We hope this clarification addresses your question, and we welcome any further discussion.

W2: The proposed SDDBMs framework recovers DDBMs, but the authors provided no theoretical derivations on the formulation of SDDBMs on VE, VP process and no empirical evidence on whether SDDBMs could improve original DDBMs.

Thank you for raising this concern. In fact, the soft VP and VE formulations can be derived directly from Eq. (14) by choosing μ=0\mu=0, θt=gt2/2\theta_t=g_t^2/2 for VP, and μ=0\mu=0, θt=0\theta_t=0 for VE, respectively. With the general parameterization x^=αx0+βx⋆+γμ\hat{x}=\alpha x_0+\beta x^\star+\gamma\mu, the resulting soft VP and VE processes are given by

dxt=[−12gt2xt+gt2e−θˉt:Tσ^2+σˉt:T2(x^−e−θˉt:Txt)]dt+gtdwtdx_{t}=\left[-\frac{1}{2}g_{t}^{2}x_{t}+g_{t}^{2}\frac{e^{-\bar{\theta}_{t:T}}}{\hat{\sigma}^{2}+\bar{\sigma}_{t:T}^{2}}\left(\hat{x}-e^{-\bar{\theta}_{t:T}}x_{t}\right)\right]dt+g_{t}dw_{t}

and

g_t^2 \frac{\hat{x}-x_t} {\hat{\sigma}^2+\bar{\sigma}_{t:T}^2}dt + g_tdw_t,$$ respectively. Furthermore, following the $x_0$-free reformulation in Section 3.4, the corresponding $x_0$-free forward SDEs can be expressed as: $$dx_t=\Bigg[g_t^2\left(\frac{e^{-\bar{\theta}_{t:T}}(\alpha-e^{-\bar{\theta}_{0:T}})}{\phi_t}-\frac12\right)x_t+\frac{g_t^2\beta e^{-\bar{\theta}_{0:T}}}{\phi_t}x^\star\Bigg]dt+\eta_t d\bar w_t$$ and $$dx_t= \Bigg[ \frac{g_t^2}{\phi_t} \left((\alpha-1)x_t+\beta x^\star\right) \Bigg]dt+\eta_t d\bar w_t,$$ respectively, where $\phi_t$ and $\eta_t$ are defined in Eqs.(18) and (21), with the corresponding VP and VE parameter settings. When $\hat{\sigma}^2\to0$, $\alpha=0$, and $\beta=1$, we have $\eta_t\to g_t$, and these two processes recover DDBM-VP and DDBM-VE in the hard-endpoint limit, respectively. Regarding the empirical evidence, due to time constraints, we have so far completed the experimental evaluation of SDDBMs-VE. As shown in Table 1, SDDBMs-VE substantially outperforms both DDBM-VE and UniDB-VE across all four evaluation metrics on the image deraining task. Moreover, it also achieves better performance than DDBM-VP and UniDB-VP across all four metrics. These results provide direct empirical evidence that SDDBMs can improve upon the original DDBMs. __Table 1. Comparison of UniDB-VP and UniDB-VE with the original DDBM-VP and DDBM-VE on image deraining.__ | Method| **PSNR ↑** | **SSIM ↑** | **LPIPS ↓** | **FID ↓** | |:---:|:---:|:---:|:---:|:---:| |DDBMs-VP|29.58|0.828|0.113|35.46| |UniDB-VP| 30.11 | 0.8414 | 0.102 | 33.17 | |DDBMs-VE|29.34|0.7654|0.185|43.22| |UniDB-VE| 29.46 | 0.7671 | 0.185| 42.57 | |SDDBMs-VE|**31.17**|**0.8655**|**0.086**|**26.19**| We are also working on the experimental evaluation of SDDBMs-VP and will report the corresponding results and incorporate them into the revised manuscript once the experiments are completed.

Q1: The qualitative visual comparisons with baselines (GOUB and UniDB) are somewhat limited. The results among GOUB, UniDB, and the proposed SDDBMs should be compared in Fig.3. More qualitative visual comparison examples would better emphasize the strengths of SDDBMs. Moreover, some examples in Fig.3 and Fig.8 are redundant. The appendix should provide additional diverse visual comparisons.

Thank you for this helpful suggestion. We have revised Fig.3 to directly compare GOUB, UniDB, and SDDBMs on the same examples for image deraining, inpainting, and super-resolution.

We have also revised Figs.7 and 8 in the appendix to include more diverse qualitative comparisons and eliminate the overlap with Fig.3. These revisions provide a more comprehensive visual comparison between SDDBMs and the baselines.

Q2: The Number of Function Evaluations (NFEs) or sampling steps for their SDDBMs and baselines are missed. Could the authors clarify the NFEs used during inference?

We apologize for omitting the number of sampling steps. Following previous works, we use 100 sampling steps (T=100T=100) for sddbms during inference. We have now includes the setting in Appendix G of the revised manuscript.

Q3: Proposition 3.1 would benefit from a rewriting. As it is, the main idea is to show the singularity of diffusion bridges. I could understand the injection (mT(x0)≠mT(y0)m_T (\mathbf{x}_0) \neq m_T (\mathbf{y}_0) for any x0≠y0\mathbf{x}_0 \neq \mathbf{y}_0) leads to the conclusion. However, in this version of Proposition 3.1, neither singularity nor the relation to diffusion bridges are emphasized.

Thank you for your helpful suggestion. We have appropriately rewritten Proposition 3.1 in the revised manuscript to explicitly emphasize the terminal singularities of diffusion bridges. Specifically, we first show that bounded drift coefficients yield a non-degenerate Gaussian terminal distribution that retains dependence on the initial state. We then prove that either complete source forgetting or terminal covariance collapse necessarily requires the linear drift coefficient to become unbounded near the terminal time. Since hard endpoint bridges simultaneously require complete source forgetting and terminal covariance collapse, they necessarily exhibit drift singularities at the terminal boundary. The corresponding proofs have also been updated in Appendix E.1.

Q4:Lack of quantitative evaluation for the ablation on hyperparameter β\beta. Although the authors have claimed that "different β\beta values lead to comparable performance", the specific results should be also provided to validate the claim, which would be more convincing.

Thanks for the helpful suggestion. we have conducted an additional quantitative ablation study on the hyperparameter β\beta for image deraining task. The results are presented in Table2.

Table 2. Effect of hyper-parameter β\beta on image deraining.

β\betaPSNR ↑SSIM ↑LPIPS ↓FID ↓
β=0.80\beta=0.8032.270.9060.04316.39
β=0.95\beta=0.9532.380.9060.04315.68
β=\beta=default32.310.9060.04216.48
β=1.30\beta=1.3032.010.90410.04518.19

Here, the "default" setting refers to a specific choice of β\beta and γ\gamma, with μ=x⋆\mu=x^\star, that makes bT+cT=1b_T+c_T=1, rather than the β\beta value used by default in our experiments.

The results show that the performance remains relatively consistent across different values of β\beta, indicating that SDDBMs are not highly sentitve to hyperparameter within the proper range (compared to σ\sigma and α\alpha). In general, values close to or slightly blew the default setting tend to yield better performance, while larger values lead to a degradation in performance. In our experiments, we set β=0.95\beta=0.95 for image deraining and adopt the default setting for both image inpainting and super-resolution.

Reviewer_Cx4wOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Soft Denoising Diffusion Bridge Models to address terminal singularities in hard-constrained diffusion bridges. The authors characterize the relationship between endpoint conditioning and singular drift coefficients, and introduce a soft Gaussian terminal constrain. The framework unifies several existing diffusion bridge models, including DDBMs, GOUB, and UniDB. Experiments demonstrate better restoration quality and numerical stability.

Critical strengths

1.The paper provides a probabilistic analysis of terminal singularities in diffusion bridges and introduces soft Gaussian terminal constraints 2. The framework provides a unified interpretation of DDBMs, GOUB, and UniDB. Experiments on three image restoration tasks show consistent improvements over existing methods.

Critical weaknesses
  1. UniDB handled terminal singularities via finite terminal penalties. While SDDBM offers a probabilistic interpretation and a more general parameterization, the practical benefit should be clearly isolated.
  2. The implementation relies on approximate terminal initialization with an empirically tuned score correction factor. Given the substantial performance gains brought by this correction, it is difficult to attribute the reported improvements exclusively to the proposed soft terminal constraints.
Questions for authors
  1. What are the advantages of SDDBMs over UniDB beyond the probabilistic reinterpretation and additional parameter flexibility?
  2. How does the approximate terminal initialization affect the theoretical guarantees of soft bridge formulation?
  3. Please provide a controlled comparison without the empirical score correction factor.
8 comments

W1: UniDB handled terminal singularities via finite terminal penalties. While SDDBM offers a probabilistic interpretation and a more general parameterization, the practical benefit should be clearly isolated.

Thank you for raising this point. Although UniDB can theoretically alleviate terminal singularities through a finite terminal penalty, its treatment of terminal stability remains limited in practice.

Our theoretical and experimental analyses identify σ\sigma and α\alpha as two key parameters governing terminal constraints and numerical stability. However, UniDB has two limitations regarding these parameters. First, the terminal variance σ2\sigma^2 in UniDB is determined by its terminal penalty coefficient γ\gamma (denoted by κ\kappa in our manuscript), and the use of a large penalty coefficient to enforce strong terminal matching results in a small σ\sigma. Second, UniDB adopts a fixed value of α=0\alpha=0, rather than allowing α\alpha to be adjusted.

These two limitations lead to two practical issues:

(1) Numerical instability near the terminal boundary. Although the finite terminal penalty in UniDB prevents the analytical drift correction from becoming singular, it does not completely eliminate numerical instability during sampling. Specifically, the coefficient of the drift correction at t=Tt=T is proportional to γ\gamma. A large penalty coefficient, corresponding to a small terminal variance, can therefore make the drift sensitive to small deviations from the target near the terminal time. Moreover, in the official UniDB-GOU implementation, the noise scale σˉt′\bar{\sigma}'_t used to convert the network prediction into a score vanishes at t=Tt=T, making the standard score update undefined at this boundary. Consequently, UniDB still requires special boundary treatment, such as starting its regular score-based reverse updates at t=T−1t=T-1.

(2) Limited flexibility in parameter selection. In UniDB, the terminal variance is controlled through the penalty coefficient γ\gamma, while α\alpha remains fixed at zero. Consequently, these two parameters cannot be flexibly and independently adjusted to balance numerical stability and restoration performance. As demonstrated by our experiments below, the parameter settings corresponding to UniDB are not necessarily optimal, and more suitable choices of σ\sigma and α\alpha can yield better restoration results.

In contrast, SDDBMs explicitly introduce σ\sigma and α\alpha as flexible parameters. Their values can be selected within the admissible parameter space to avoid numerical instability near the terminal boundary while achieving better restoration performance. This provides a practical advantage over the restricted parameter settings of UniDB.

To further demonstrate these advantages, we provide the following theoretical and experimental analyses of σ\sigma and α\alpha.

Effect of σ\sigma. We first study the effect of the terminal variance parameter σ2\sigma^2, which controls the strength of the prescribed Gaussian terminal constraint in SDDBMs. As σ\sigma approaches zero, the conditional score becomes singular, leading to a terminal singularity at t=Tt=T. To examine how this affects the sampling dynamics in practice, we track three quantities at a fixed pixel location throughout reverse sampling: the value of xt\mathbf{x}_{t}, the drift term in Eq. (17), and the network prediction ϵθ(xt,x⋆,t)\boldsymbol{\epsilon}_{\theta}(\mathbf{x}_{t}, \mathbf{x}^{\star}, t). As shown in Figure 2(a), all three quantities exhibit large fluctuations when σ=0.001\sigma=0.001. These fluctuations are substantially smaller at σ=0.05\sigma=0.05 and σ=0.1\sigma=0.1, indicating more stable sampling behavior. We next examine how this sampling behavior relates to restoration performance. Figure 3(a) shows that very small values of σ\sigma yield low PSNR on image deraining, consistent with the large fluctuations observed during sampling. Once σ\sigma reaches 0.050.05, performance becomes relatively stable, with PSNR peaking in the range [0.08,0.11][0.08, 0.11] before declining as σ\sigma increases further. The value σ=0.05\sigma=0.05 also provides a natural point of comparison with GOUB and UniDB. With T=100T=100, both implementations begin their regular score-based reverse updates at t=99t=99, where the bridge noise standard deviation is approximately 0.050.05. In this sense, their practical sampling configurations can be viewed as approximate special cases of SDDBMs with σ≈0.05\sigma\approx 0.05. The flexibility to choose σ\sigma allows SDDBMs to move beyond this baseline setting and attain the higher PSNR observed in the range [0.08,0.11][0.08, 0.11]. In addition to its effect on restoration performance, σ\sigma also influences training stability. As shown in Figure 5(a), the training loss fluctuates considerably when σ\sigma is very small, whereas moderate values of σ\sigma yield lower and more stable training losses.

Effect of α\alpha. We next investigate the effect of α\alpha. As derived in Appendix D.2, the lower bound for admissible values of α\alpha is

αmin⁡=−e−θˉ0:Tσ^2σˉ0:T2.\alpha_{\min}=-\frac{e^{-\bar{\theta}_{0:T}}\hat{\sigma}^{2}}{\bar{\sigma}_{0:T}^{2}}.

At α=αmin⁡\alpha=\alpha_{\min}, the denominator φT\varphi_T in the drift coefficients of Eq. (17) vanishes, which also leads to a terminal singularity. Choosing α\alpha close to αmin⁡\alpha_{\min} can therefore cause numerical instability. To assess how this instability affects reverse sampling in practice, we compare the pixel trajectory, the drift term, and the network prediction for values of α\alpha close to and farther from αmin⁡\alpha_{\min}, as shown in Figure 2(b). We observe the same pattern as for very small σ\sigma. All three quantities exhibit large fluctuations when α\alpha is close to αmin⁡\alpha_{\min} and are more stable for values farther from this boundary.

We also examine how this sampling behavior relates to restoration performance. As shown in Figure 3(b), PSNR drops as α\alpha approaches αmin⁡\alpha_{\min}, while small positive values of α\alpha yield higher PSNR than α=0\alpha=0. As discussed in Section 4, both GOUB and UniDB correspond to α=0\alpha=0 in our framework. As with σ\sigma, allowing α\alpha to vary beyond the baseline setting improves restoration performance in our experiments. Furthermore, we observe that α\alpha also affects the training loss, similar to σ\sigma (Figure 5 (b)).

We have also provided a detailed analysis in Section 5 of the revised manuscript. We hope these revisions and clarifications address your concerns.

W2: The implementation relies on approximate terminal initialization with an empirically tuned score correction factor. Given the substantial performance gains brought by this correction, it is difficult to attribute the reported improvements exclusively to the proposed soft terminal constraints.

Thank you for raising these concerns regarding the approximate terminal initialization and the empirically tuned score correction factor. We have addressed these two issues separately:

(1) Terminal initialization. In our response to Q2, we provide a detailed explanation of why the approximate terminal initialization is reasonable and why it does not affect the theoretical derivations of SDDBMs.

(2) Empirical score correction factor. In our response to Q3, we provide additional experimental results with ζ=1\zeta=1, demonstrating that SDDBMs still achieve the best overall performance even without the empirical score correction factor. Furthermore, to make the main contribution of our work clearer, we have removed all discussions of ζ\zeta and the corresponding experiments from the revised manuscript.

We emphasize that, after removing ζ\zeta, our experimental implementation fully aligns with the theoretical formulation of SDDBMs. Importantly, SDDBMs still outperform the baselines, demonstrating the advantages of the proposed framework independently of the empirical score correction factor.

Reviewer_FZyuOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Soft Denoising Diffusion Bridge Models, which replace hard endpoint constraints with a non-degenerate Gaussian terminal distribution to address terminal singularities. The framework derives closed-form terminal reweighting, forward marginals, and an x0x_0-free reformulation of the dynamics, while recovering DDBM, GOUB, and UniDB as special or limiting cases. Experiments on image deraining, super-resolution, and inpainting report improvements over existing diffusion bridge methods.

Critical strengths
  1. The paper presents a unified probabilistic formulation of diffusion bridges with analytically tractable soft terminal constraints.
Critical weaknesses
  1. Figure 1 illustrates a transformation between different image domains, and the conclusion claims performance across image-to-image translation and restoration tasks. However, all experiments are image restoration tasks. The experimental scope does not substantiate the broader image-translation claims.
  2. The headline results in Table 2 correspond to the corrected samplers in Table 3. This correction has a substantial effect: on super-resolution, the SDE result changes from 22.81 to 27.22 dB in PSNR and from 0.366 to 0.136 in LPIPS. Without the correction, the proposed SDE underperforms GOUB and UniDB on this task. It remains unclear whether comparable tuning was afforded to the baselines, and how much of the improvement should be attributed to the proposed probabilistic formulation. The reported improvements are not sufficiently separated from an additional empirical sampling correction.
Questions for authors
  1. What is the intended takeaway of Figure 2? As presented, the figure is confounded: it measures performance under varying hyperparameters, which does not directly establish numerical stability.
  2. What is Figure 4 meant to demonstrate? It shows that moderate values of sigmasigma yield lower training losses, but all displayed settings appear to converge. As such, the figure does not convincingly connect the observed training behavior to improved numerical conditioning at the terminal boundary.
  3. Table 3 shows a substantial difference between the standard SDE and SDE-zetazeta, especially on super-resolution. How were the task-specific zetazeta values selected, and were GOUB and UniDB given an equivalent tuning budget?
  4. Can the authors evaluate on cross-domain translation benchmarks, such as Edges→Handbags and DIODE-Outdoor, as used in DDBM?
  5. Where should the visual improvements in the second deraining case of Figure 3 be observed? At the displayed resolution, the proposed method appears visually comparable to the baseline. Moreover, the figure omits the degraded inputs, making it difficult to assess the severity of the degradation and the extent of the improvement.
  6. How is the terminal distribution handled when x0x_0 is unavailable? Please explicitly distinguish x0x_0-independent dynamics from a terminal distribution that still depends on x0x_0.
  7. Why is the Related Work section relegated to the appendix rather than incorporated into the main manuscript? It provides essential context for situating the contributions and would be better placed in the main body.
8 comments

W1&Q4: Figure 1 illustrates a transformation between different image domains, and the conclusion claims performance across image-to-image translation and restoration tasks. However, all experiments are image restoration tasks. The experimental scope does not substantiate the broader image-translation claims. Can the authors evaluate on cross-domain translation benchmarks, such as Edges→Handbags and DIODE-Outdoor, as used in DDBM?

Thank you for raising this concern. We have conducted additional experiments on the Edges→Handbags dataset, comparing SDDBMs with GOUB and UniDB under the same experimental settings as those used in G2 (we set batch size=128 and train 400000 steps for all the models). As shown in Table 1, the results demonstrate that SDDBMs also outperform the baselines on this image-to-image translation benchmark. The corresponding results are also presented in Table 4 of the revised manuscript.

Table 1. Comparison of SDDBMs, UniDB, and GOUB on the Edges→Handbags image-to-image translation benchmark.

MethodsPSNR ↑SSIM ↑LPIPS ↓FID ↓
SDDBMs17.020.5930.15265.04
UniDB16.740.5850.16266.51
GOUB16.740.5670.16666.81

Due to time constraints, we plan to further include comparisons with DDBMs and additional experiments on the DIODE-Outdoor dataset in a subsequent revision.

W2: The headline results in Table 2 correspond to the corrected samplers in Table 3. This correction has a substantial effect: on super-resolution, the SDE result changes from 22.81 to 27.22 dB in PSNR and from 0.366 to 0.136 in LPIPS. Without the correction, the proposed SDE underperforms GOUB and UniDB on this task. It remains unclear whether comparable tuning was afforded to the baselines, and how much of the improvement should be attributed to the proposed probabilistic formulation. The reported improvements are not sufficiently separated from an additional empirical sampling correction.

Thank you for this insightful comment. We apologize that the introduction of the empirical correction factor ζ\zeta may have obscured the main contribution of SDDBMs. To clarify the performance and contribution of SDDBMs independently of this factor, we have removed all discussions of ζ\zeta from the revised manuscript and rerun the experiments without this empirical correction (ζ=1\zeta=1).

To address your concerns in detail, we clarify three points below: (1) the performance of SDDBMs without the empirical correction factor; (2) the differences between Table 1 in this response and Table 3 of the original manuscript, which arise from different training checkpoints for the super-resolution task, together with the reasons for selecting these checkpoints; and (3) the hyperparameter tuning of the baseline methods.

(1) Performance without the empirical correction factor ζ\zeta.

We first provide additional comparisons in Table 2, including the performance of SDDBMs without the empirical correction factor (ζ=1\zeta=1) and the corresponding results with the correction. Even with ζ=1\zeta=1, SDDBMs outperform GOB and UniDB on all four metrics across the three tasks when comparing the corresponding samplers. These results demonstrate that performance improvements are already present without the ζ\zeta adjustment.

Table 2. Comparison of SDDBMs with the published baseline results. SDE and Mean-ODE denote the uncorrected SDDBMs samplers with ζ=1\zeta=1.

METHODImage DerainingImage Super-ResolutionImage Inpainting
PSNR ↑SSIM ↑LPIPS ↓FID ↓PSNR ↑SSIM ↑LPIPS ↓FID ↓PSNR ↑SSIM ↑LPIPS ↓FID ↓
SDE32.380.90690.04215.6826.280.72000.17816.0729.620.90840.0343.86
Mean-ODE34.960.94500.06928.4428.800.81410.31420.3032.000.94040.05111.84
SDE-ζ32.940.91830.03814.2127.220.75220.13613.7030.300.92010.0323.48
ODE-ζ35.060.94640.06625.7528.900.81640.30919.9532.220.93990.05111.87
GOUB-SDE31.960.90280.04618.1426.890.74780.22020.8528.980.90670.0374.30
GOUB-ODE34.560.94140.07732.8328.500.80700.32822.1431.390.93920.05212.24
UniDB-SDE32.050.90360.04517.6525.460.68560.17916.2129.200.90770.0364.08
UniDB-ODE34.680.94260.07431.1628.640.80720.32322.3231.670.93950.05211.98

(2) Explanation of the differences between Table 1 and the original Table 3.

For super-resolution with the uncorrected SDE sampler, Table 2 reports the results obtained using a network trained for 20,000 steps. The corresponding results in Table 3 of the original manuscript were obtained using a network trained for 600,000 steps. The SDE-ζ\zeta entry in Table 2 retains the results from the network trained for 600,000 steps, so these two super-resolution entries use different training checkpoints.

We made this addition because we observed substantial fluctuations in the super-resolution performance of the uncorrected SDE sampler as training continued beyond 20,000 steps. Similar behavior also occurs in GOUB and UniDB. To examine this behavior, we separately evaluated the SDE samplers of all three methods at different training checkpoints, as shown in Table 3.

Table 3. Super-resolution PSNR of GOUB, UniDB, and SDDBMs using their SDE samplers at different training checkpoints. For SDDBMs, ζ=1\zeta=1.

Training stepsGOUBUniDBSDDBMs
2000025.2425.6026.28
4000024.5922.8123.39
6000024.1324.1625.46
8000023.2321.2823.69
10000023.0922.7924.93
12000022.2124.2726.39
14000020.8524.3524.11
16000025.5523.2527.03
18000023.4623.5223.83
20000025.8624.1622.95

These results show that additional training does not consistently improve SDE sampling performance. For SDDBMs, the network trained for 20,000 steps achieves a PSNR of 26.28, while some later checkpoints achieve higher values, including 27.03 at 160,000 steps. However, performance fluctuates considerably across checkpoints. We therefore present the 20,000-step result as a well-performing early checkpoint of the uncorrected sampler, without claiming that it is the best result achievable during training.

In our experiments, we found that an appropriate choice of ζ\zeta helped alleviate these fluctuations, which motivated the use of the empirical correction. For super-resolution in the original Table 3, we used the network trained for 600,000 steps because the corrected sampler achieved its best performance at that checkpoint. We also evaluated the uncorrected sampler using the same network, providing a comparison at the same training checkpoint. The additional results presented here complement that comparison by showing the better performance that the uncorrected sampler can achieve earlier in training.

(3) the hyperparameter tuning of the baseline methods. The GOUB and UniDB results in Table 1 are the best results reported by the authors in their respective papers, rather than results from our own reproductions. We use these published results directly, without any additional hyperparameter tuning.

We apologize for not clearly explaining this checkpoint choice and for not reporting these stronger results for the uncorrected sampler. We hope that the additional experiments and clarification address your concern.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.