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

Hai Huang; Hanting Wang; Shengpeng Ji; Shulei Wang; Tao Jin; Wang Lin; Zhou Zhao

REVIEW 3 major objections 4 minor 4 cited by

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

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Image restoration bridges can be driven by pretrained diffusion models through a single transition equation, eliminating the need to train a new bridge for each degradation type.

desk verdict A practically useful trick for plugging pretrained diffusion models into restoration bridges, but the central proof is in pixel space while the experiments run in latent space, and the schedules are tuned on test data. read the letter →

arxiv 2505.24406 v1 pith:2NL2MKXH submitted 2025-05-30 cs.CV

Hanting Wang , Tao Jin , Wang Lin , Shulei Wang , Hai Huang , Shengpeng Ji , Zhou Zhao This is my paper

classification cs.CV

keywords imagerestorationdiffusionbridgemodelstransitionequationcriticaltimesteppretrainedpriorsIR-SDEGOUBStable

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that an image restoration bridge — a diffusion process that runs from a degraded image to a clear one — can be solved using a pretrained generative diffusion model, without training a new bridge for each degradation type. The central claim is Proposition 3.1: when two diffusion processes share the same endpoint distribution, their states can be moved into each other by the transition equation $x^*_j = \alpha\,x^\circ_i + \beta\,x_0 + \gamma + \sigma\epsilon$, and at a chosen critical timestep the coefficient $\beta$ of the unknown clean image vanishes, so a pretrained denoiser can estimate $x_0$ directly. If the claim holds, one conditional generative model (Stable Diffusion v1-5 with ControlNet) handles deraining, dehazing, desnowing, raindrop removal, low-light enhancement, and image inpainting, with less training and better generalization than bridges trained from scratch. The paper reports experiments on all six tasks with consistent gains, especially in FID and in generalization to scenes outside the training distribution.

What carries the argument

The load-bearing object is the transition equation of Proposition 3.1, $x^*_j = \alpha\,x^\circ_i + \beta\,x_0 + \gamma + \sigma\epsilon$, with $\alpha = \sqrt{((\sigma^*_j)^2 - \sigma^2)/(\sigma^\circ_i)^2}$, $\beta = f^*_j - \alpha f^\circ_i$, and $\gamma = b^*_j - \alpha b^\circ_i$, valid when $\max\{0,\, [(\sigma^*_j)^2 - (f^*_j\sigma^\circ_i/f^\circ_i)^2]\} \le \sigma^2 \le (\sigma^*_j)^2$. Its companion concept is the critical timestep $\tilde{t}_i$, the smallest $j$ at which the coefficient $\beta$ can be driven to zero; choosing $j$ above it with $\sigma$ at its minimum gives the forward transition (adding noise without reference to $x_0$), while reversing the roles with $\sigma = 0$ gives a deterministic reverse transition back toward the bridge. This pair of moves lets a pretrained denoiser act on restoration-bridge states as if they were its own, which is the mechanism that makes restoration possible without training a bridge from scratch.

What would settle it

Compute the empirically transformed state's marginal under the forward transition for a range of bridge parameters $\lambda$ and degradation severities, compare it with the pretrained model's marginal $N(x; \sqrt{\bar\alpha_j}\,x_0,\, (1-\bar\alpha_j)I)$ at the same timestep, and measure the denoiser's prediction error $\|\hat{x}_0 - x_0\|$; if the error tracks the distributional gap and grows sharply as the gap widens, the framework's reliance on out-of-distribution denoising stands exposed as its binding constraint.

Watch

Extended reading notes

Core claim

The paper's discovery is a transition equation that bridges any two linear Gaussian diffusion paths with the same endpoint distribution. Reparameterizing a generative diffusion model and an image restoration bridge both as $x_t = f_t x_0 + b_t + \sigma_t \epsilon$, the equation $x^*_j = \alpha\,x^\circ_i + \beta\,x_0 + \gamma + \sigma\epsilon$ transfers a bridge state into a generative state and back, with the coefficients fixed by matching the two conditional distributions under conditional independence given $x_0$. The boundary analysis shows that at the critical timestep $j = \tilde{t}_i$, where $(f^\circ_i/f^*_j)^2 = (\sigma^*_j/\sigma^\circ_i)^2$, choosing the minimal noise $\sigma$ makes $\beta = 0$, so the forward transition no longer depends on the unknown clean image $x_0$; the pretrained denoiser then estimates $\hat{x}_0$, and a deterministic reverse transition ($\sigma = 0$, $\beta \ge 0$) returns the estimate to the bridge's previous timestep. Iterating these transitions lets a pretrained conditional generative model drive the restoration bridge end to end, which the paper argues removes per-degradation training and improves robustness and generalization because the generative prior was trained on large-scale image data.

Load-bearing premise

The framework rests on the pretrained denoiser estimating $x_0$ accurately from forward-transitioned states that do not follow the generative model's own marginal distribution at the chosen timestep — a distributional mismatch the paper does not analyze, and one that would degrade or collapse the restoration loop if the denoiser were sensitive to it.

Editorial extensions

If this is right

One conditional generative model, Stable Diffusion v1-5 with ControlNet, drives all six tested restoration tasks through the transition equation and outperforms most per-degradation bridges trained from scratch, with the largest gains in FID.
Training cost drops substantially: the ControlNet used by IRBridge trains in about one day on an Nvidia 3090, compared with about 2.5 days to train GOUB from scratch.
Generalization to unseen domains improves: an inpainting model trained only on facial images still restores indoor and outdoor scenes, and IRBridge leads on the no-reference metrics MUSIQ, BRISQUE, and NIQE on real-world rain, haze, and snow images.
Inference can be shortened without much loss: cutting the process from 100 to 25 steps keeps PSNR and SSIM nearly unchanged while reducing runtime from 14.2 seconds to 3.4 seconds per image.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The paper plots critical timesteps for DDPM and Rectified Flow and finds them in the usable range, but only tests Stable Diffusion; running IRBridge on a Rectified Flow or DDPM denoiser would directly confirm whether the transition equation generalizes across generative paradigms.
Since Section 5.4 concedes that the proposed reverse transition is not the optimal reverse state, swapping in the original bridge's reverse SDE is a clean test of whether the deterministic reverse path is the performance bottleneck, and could raise PSNR toward the bridge's own ceiling.
The per-task timestep schedules are empirically tuned; the critical-timestep curves suggest an automatic schedule that keeps every forward-transitioned state inside the denoiser's trained range, which would remove the task-specific tuning the paper lists as a limitation.
The transition equation itself is agnostic to images, so the same mechanism could carry pretrained priors into other bridge problems with a shared endpoint distribution, such as image translation or style transfer, where the paper's impact statement already flags misuse risk.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Bottom line: the paper has a nice practical trick, but the central theorem is proven in the wrong space for the experiments. Worth reading, but don't take the theoretical claims at face value.

What's new: the transition equation itself is, as the authors admit, a reparameterization of two linear Gaussian paths. The contribution is the recipe: use the critical timestep to set the x0 coefficient to zero, so you can map a restoration bridge state into a pretrained diffusion state without knowing x0, denoise, and map back. That's a clean idea, and the six-task experiments show that with a ControlNet you can replace a from-scratch bridge model with a pretrained Stable Diffusion plus a light adapter, with big gains in FID/LPIPS. The generalization result on inpainting non-facial scenes is genuinely encouraging.

The soft spots:

1. The latent-space gap is the big one. Proposition 3.1 is derived for two linear-Gaussian processes in the same coordinate space. The experiments use Stable Diffusion, which runs in VAE latent space, while IR-SDE and GOUB are defined on pixels. The paper never says whether the forward transition is computed in pixel space and then VAE-encoded, or whether the whole bridge is run on latents. Neither option is covered by the proof: the VAE encoder is nonlinear, so the transformed state is not a valid DDPM latent, and if you re-derive the bridge in latent space, the coefficients in Table 2 no longer apply. The empirical results may still hold because ControlNet is flexible, but the paper overstates what Proposition 3.1 establishes.

2. The schedules are tuned on test data. Appendix D says they picked among eight predefined timestep schedules using a mini-batch of test images, then fine-tuned further. No error bars are reported. That leaves the quantitative comparison on shaky ground, especially since the text claims "outperforms" bridge models while PSNR is frequently lower than the from-scratch baselines; the gains are mostly in perceptual metrics.

3. Minor: the "critical timestep" analysis is useful, but the paper doesn't give the final schedules numerically, which makes reproduction harder.

Who should read it: anyone working on turning pretrained diffusion models into restoration backbones. The framework is likely to be built on even if the current theory is incomplete. I wouldn't cite it as a theoretical foundation, but I'd cite it as a practical recipe once the latent-space question is addressed. A serious referee should ask for the code, exact schedules, error bars, and either a re-derivation of the transition for latents or a clear statement that the theory is pixel-space only and the latent case is heuristic. I'd accept it for review, but with heavy revision expected.

Referee Report

3 major / 4 minor

Summary. The paper proposes IRBridge, a framework that integrates pretrained generative diffusion models (Stable Diffusion v1.5) into image restoration bridge models (IR-SDE and GOUB) without training a restoration bridge from scratch. The key contribution is a transition equation (Proposition 3.1, Eq. 7) that maps states of a bridge process to states of a generative diffusion process sharing the same endpoint (clean-image) distribution. Under a linear-Gaussian reparameterization, the equation gives coefficients that, at a critical timestep, make the mapping independent of the clean image x0, enabling a 'forward transition' from bridge to generative states; a 'reverse transition' maps the estimated x0 back to the bridge trajectory. Experiments on six restoration tasks report PSNR/SSIM/LPIPS/FID comparisons against bridges trained from scratch and against other prior-based methods, and additional real-world generalization results.

Significance. The central algebraic idea—bridging two diffusion processes by matching the coefficients of their linear-Gaussian conditionals—is clean, and Appendix A's derivation is correct under its stated assumptions. If the coordinate-space gap with latent Stable Diffusion is resolved, the framework could offer a practical way to leverage large pretrained generative priors in restoration bridges, and the paper demonstrates a wide range of tasks and a real-world generalization study. The framework's flexibility in choosing bridge and generative models is a strength. However, the theoretical bridge currently overstates what is proven for the actual experimental setup, and the empirical evaluation is undermined by test-set schedule selection and ambiguous evaluation references; these issues presently limit the significance of the claimed improvements.

major comments (3)

[Section 3.1 / Appendix A (Eq. 7) vs. Section 4.1] Proposition 3.1 is proved for two linear-Gaussian processes in the same coordinate space, but the experiments pair pixel-space restoration bridges (IR-SDE, Eq. 3–4; GOUB, Eq. 5–6) with the latent-space Stable Diffusion v1.5. The paper does not state whether the forward transition is computed on pixels and then VAE-encoded, or computed in latent space after encoding the bridge state. In the first case, the VAE encoder is nonlinear, so E(α x°_i + γ + σ ε) does not follow the latent-space Gaussian form of Eq. (1); in the second case, the pixel-space coefficients f°_i and σ°_i used in Eq. (7) do not describe the conditional distribution of the encoded latent. Consequently, Proposition 3.1 does not, as written, establish that the transformed state is a valid forward state of the pretrained generative model. This is load-bearing because the framework's premise is that the pretrained denoiser can be applied to the transformed state. The empirical results may still hold if the denoiser/ControlNet tolerate out-of-distribution latents, but the theoretical bridge is not established. Please either (a) run the restoration bridge entirely in latent space so that the linear-Gaussian assumption holds, or (b) provide a corrected analysis that explicitly accounts for the VAE map (e.g., a derived bound on the distribution shift, or at least an empirical validation that the transformed latents are close to the training marginal).
[Appendix D / Section 5.1] The per-task timestep schedules used in Table 1 are selected after evaluating eight predefined schedules on mini-batches of four images, and the paper does not state whether these mini-batches were drawn from the training or test sets. The final schedules are also not reported numerically (Figure 4 shows only qualitative curves, and Appendix D states that 'slight discrepancies' remain after fine-tuning). This makes the comparison susceptible to selection bias and prevents exact reproduction. The paper itself acknowledges that the schedules are 'not the optimal choice' (Section 5.1), but the extent of tuning is not transparent. Please specify the data split used for schedule selection, report the exact schedule values per task, and provide results with multiple seeds or error bars to support the robustness claims.
[Section 4.2] The sentence 'To eliminate potential interference introduced by the VAE employed by the pretrained stable diffusion model, these metrics are computed using label images decoded by the same VAE' is ambiguous and potentially biasing. If IRBridge outputs are compared against VAE-decoded ground truth while the baselines are compared against the original ground truth, the comparison is not apples-to-apples; if instead all methods are compared against VAE-decoded labels, the pixel-space baselines are evaluated on an unfamiliar reference. This ambiguity directly affects the interpretation of Table 1. Please clarify which reference was used for each method and report at least one set of metrics computed on the original ground-truth images.

minor comments (4)

[Section 3.2 / Figure 2] The phrase 'Critial Timestep' is misspelled (should be 'Critical Timestep'), and the same typo appears in Figure 2's caption as 'Critial timsteps'.
[Figure 2 caption] The caption says 'Stabel-Diffusion v1-5'; the intended spelling is 'Stable Diffusion'.
[Appendix C] The section refers to 'Recited Flow'; the correct name is 'Rectified Flow'.
[Appendix D] The text says 'Figure (12,13,14,15,16,16,17)', listing Figure 16 twice; the duplicate should be removed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Proposition 3.1 is an explicit algebraic reparameterization with no fitted inputs; the caveats are benchmark-selection and coordinate-mismatch issues, not circular reductions.

full rationale

The central derivation is Proposition 3.1, proved in Appendix A by substituting the two linear-Gaussian reparameterizations into the assumed transition equation and solving the coefficient equations (f*_j = alpha*f_deg_i + beta, b*_j = alpha*b_deg_i + gamma, (sigma*_j)^2 = (alpha*sigma_deg_i)^2 + sigma^2). This is an exact algebraic identity, not a fit: no parameter of the transition equation is learned from or tuned to the restoration results, and the paper explicitly states that Eq. 7 is a reparameterization. The boundary conditions, critical timestep, and forward/reverse transitions are likewise derived from the known diffusion coefficients of the two processes, not from the experimental outputs. The empirical gains rest on external components: the pretrained Stable Diffusion v1-5 denoiser and a separately trained ControlNet, neither of which is a renamed version of the transition equation. Two caveats appear in the manuscript but do not constitute circularity. First, per-task timestep schedules were selected by evaluating predefined schedules on small batches (Appendix D) and then fine-tuned, which is a benchmark-selection and soundness concern rather than a reduction of the prediction to its inputs. Second, the paper applies a pixel-space bridge theorem to a latent diffusion backbone (Stable Diffusion) without specifying how the VAE coordinate transform is handled, which is a validity gap in the application of the theorem, not a circular step. The reference list contains several self-citations by the authors, but none is load-bearing in the derivation and none is invoked as an external uniqueness or justification theorem. Therefore no circular step is exhibited, and the derivation is self-contained relative to its stated assumptions.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No new entities are postulated. The theoretical core is a reparameterization with no fitted constants; the load-bearing tunable elements are the per-task timestep schedules and lambda, both empirically selected. The framework also depends on two unproven empirical assumptions about the pretrained denoiser and the conditioning adapter.

free parameters (3)

IR-SDE/GOUB lambda = 2 (default; 1 to 5 explored in Figure 2)
Scale of the mean-reverting noise in Equations 3 and 5; affects critical timestep locations; set by hand and not optimized.
Per-task DDPM timestep schedule = Varies by task, e.g., linear decay from j=850 or j=550 to 1; exact final schedules not tabulated
Chosen per task by running 8 predefined schedules on mini-batches of 4 test images and picking the one with stable metric curves (Appendix D); a fitted inference parameter set.
Number of inference steps = 100 (50 and 25 also evaluated)
Iteration count for the forward and reverse transition loop; user-set hyperparameter (Table 3).

assumptions (4)

domain assumption Both generative and restorative diffusion processes can be written as linear Gaussian paths x_t = f_t x0 + b_t + sigma_t epsilon
Stated in Section 2 and Appendix B; true for DDPM, IR-SDE, GOUB, rectified flow, RDDM, and DiffUIR as tabulated, but is a structural assumption on the processes.
domain assumption The two diffusion processes are conditionally independent given x0
Used in the proof of Proposition 3.1 in Appendix A to justify the reparameterization; reasonable for independent SDEs but an assumption about the coupling.
ad hoc to paper The pretrained Stable Diffusion denoiser remains accurate on forward-transformed bridge states that do not necessarily match its training marginal
The beta=0 forward transition ignores x0 and produces a state whose distribution is not shown to equal the DDPM marginal at timestep j; the method's success relies on this empirical tolerance (Section 3.2).
domain assumption ControlNet conditioning forces the estimated x0 to be consistent across iterations
Section 5.2 shows unconditional models fail because per-step x0 estimates conflict; the framework assumes a trained ControlNet provides enough consistency to avoid error accumulation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of IRBridge: Solving Image Restoration Bridge with Pre-trained Generative Diffusion Models." pith.science (2026). https://pith.science/paper/2NL2MKXH

@misc{pith2026250524406,
  author       = {Pith},
  title        = {Pith review of: IRBridge: Solving Image Restoration Bridge with Pre-trained Generative Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2NL2MKXH}},
  note         = {Machine review of arXiv:2505.24406}
}

read the original abstract

Bridge models in image restoration construct a diffusion process from degraded to clear images. However, existing methods typically require training a bridge model from scratch for each specific type of degradation, resulting in high computational costs and limited performance. This work aims to efficiently leverage pretrained generative priors within existing image restoration bridges to eliminate this requirement. The main challenge is that standard generative models are typically designed for a diffusion process that starts from pure noise, while restoration tasks begin with a low-quality image, resulting in a mismatch in the state distributions between the two processes. To address this challenge, we propose a transition equation that bridges two diffusion processes with the same endpoint distribution. Based on this, we introduce the IRBridge framework, which enables the direct utilization of generative models within image restoration bridges, offering a more flexible and adaptable approach to image restoration. Extensive experiments on six image restoration tasks demonstrate that IRBridge efficiently integrates generative priors, resulting in improved robustness and generalization performance. Code will be available at GitHub.

Figures

Figures reproduced from arXiv: 2505.24406 by the authors.

**Figure 2.** An example of critical timesteps: the generative [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 4.** Timestep schedules selected for different image [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗

Figures from the paper (22 more)

**Figure 5.** Figure 5: The impact of timestep scheduling on the model’s prediction [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

**Figure 6.** Figure 6: The variation of the forward/reverse transition [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

**Figure 7.** Figure 7: We evaluate the generalization ability of other [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]

**Figure 8.** Figure 8: Ablation of Condition for IRBridge. We visualize [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

**Figure 9.** Figure 9: Comparison of the standard generative process [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]

**Figure 10.** Figure 10: Critical timesteps across different paradigms. We cover commonly used generative models (Stable Diffusion [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]

**Figure 11.** Figure 11: Predefined timestep schedules for exploratory study on timestep scheduling strategies. [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]

**Figure 12.** Figure 12: Exploratory study on timestep scheduling strategy for the image deraining task. [PITH_FULL_IMAGE:figures/full_fig_p015_12.png]

**Figure 13.** Figure 13: Exploratory study on timestep scheduling strategy for the image dehazing task. [PITH_FULL_IMAGE:figures/full_fig_p016_13.png]

**Figure 14.** Figure 14: Exploratory study on timestep scheduling strategy for the image desnowing task. [PITH_FULL_IMAGE:figures/full_fig_p016_14.png]

**Figure 15.** Figure 15: Exploratory study on timestep scheduling strategy for the image raindrop removal task. [PITH_FULL_IMAGE:figures/full_fig_p017_15.png]

**Figure 16.** Figure 16: Exploratory study on timestep scheduling strategy for the low light enhancement task. [PITH_FULL_IMAGE:figures/full_fig_p017_16.png]

**Figure 17.** Figure 17: Exploratory study on timestep scheduling strategy for the image inpainting task. [PITH_FULL_IMAGE:figures/full_fig_p018_17.png]

**Figure 18.** Figure 18: Visualization results of IRBridge in real-world rainy images. [PITH_FULL_IMAGE:figures/full_fig_p020_18.png]

**Figure 19.** Figure 19: Visualization results of IRBridge in real-world hazy images. [PITH_FULL_IMAGE:figures/full_fig_p020_19.png]

**Figure 20.** Figure 20: Visualization results of IRBridge in real-world snowy images. [PITH_FULL_IMAGE:figures/full_fig_p020_20.png]

**Figure 21.** Figure 21: Visualization results of IRBridge in the image deraining task. [PITH_FULL_IMAGE:figures/full_fig_p021_21.png]

**Figure 22.** Figure 22: Visualization results of IRBridge in the image dehazing task. [PITH_FULL_IMAGE:figures/full_fig_p021_22.png]

**Figure 23.** Figure 23: Visualization results of IRBridge in the image desnowing task. [PITH_FULL_IMAGE:figures/full_fig_p022_23.png]

**Figure 24.** Figure 24: Visualization results of IRBridge in the image raindrop removal task. [PITH_FULL_IMAGE:figures/full_fig_p022_24.png]

**Figure 25.** Figure 25: Visualization results of IRBridge in the low light enhancement task. [PITH_FULL_IMAGE:figures/full_fig_p023_25.png]

**Figure 26.** Figure 26: Visualization results of IRBridge in the image inpainting task. [PITH_FULL_IMAGE:figures/full_fig_p023_26.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 4 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

TAP: Parameter-efficient Task-Aware Prompting for Adverse Weather Removal
cs.CV 2025-08 conditional novelty 6.0 of 10

A two-stage prompt-tuning method with low-rank and contrastive prompt enhancement claims all-in-one adverse weather removal at 2.75M parameters.
Show and Polish: Reference-Guided Identity Preservation in Face Video Restoration
cs.CV 2025-07 conditional novelty 6.0 of 10

IP-FVR restores degraded face videos with consistent identity by conditioning a video diffusion model on a reference photo of the same person.
Bridging Domain Generalization to Multimodal Domain Generalization via Unified Representations
cs.CV 2025-07 conditional novelty 6.0 of 10

By aligning category-level information across video, audio, and flow into one unified representation while keeping modality-specific details separate, this paper shows that standard domain generalization methods impro...
Open-set Cross Modal Generalization via Multimodal Unified Representation
cs.CV 2025-07 conditional novelty 5.0 of 10

The authors propose OSCMG, an open-set version of Cross Modal Generalization, and show their MICU method with masked contrastive learning and unified jigsaw puzzles outperforms prior methods.

Reference graph

Works this paper leans on

23 extracted references · 7 canonical work pages · cited by 4 Pith papers

[2]

the results listed in Table 1 are obtained with fine-tuned timestep schedules, which are inherently better than the predefined schedules used in the exploratory experiments. We argue that using small-batch samples in exploratory studies is feasible because our primary focus is on observing the evolution curves of PSNR and SSIM rather than their final valu...

work page
[6]

Li, B., Xue, K., Liu, B., and Lai, Y .-K

doi: 10.1109/TIP.2018.2867951. Li, B., Xue, K., Liu, B., and Lai, Y .-K. Bbdm: Image-to- image translation with brownian bridge diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern Recognition, pp. 1952–1961, 2023a. Li, T., Ku, M., Wei, C., and Chen, W. Dreamedit: Subject- driven image editing, 2023b. Li, W., Zhang, Q....

arXiv 2018
[8]

Flow straight and fast: Learning to generate and transfer data with rectified flow

10 IRBridge: Solving Image Restoration Bridge with Pre-trained Generative Diffusion Models Liu, X., Gong, C., and Liu, Q. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022b. Liu, Y .-F., Jaw, D.-W., Huang, S.-C., and Hwang, J.-N. Desnownet: Context-aware deep network for snow re- mova...

arXiv
[10]

Peng, X., Zheng, Z., Dai, W., Xiao, N., Li, C., Zou, J., and Xiong, H

doi: 10.1109/TPAMI.2023.3238179. Peng, X., Zheng, Z., Dai, W., Xiao, N., Li, C., Zou, J., and Xiong, H. Improving diffusion models for inverse prob- lems using optimal posterior covariance. InForty-first International Conference on Machine Learning,

arXiv 2023
[12]

GenDeg: Diffusion-based Degradation Synthesis for Generalizable All-In-One Image Restoration

Rajagopalan, S., Nair, N. G., Paranjape, J. N., and Patel, V . M. Gendeg: Diffusion-based degradation synthesis for generalizable all-in-one image restoration.arXiv preprint arXiv:2411.17687,

work page Pith review arXiv
[14]

P., Kumar, A., Er- mon, S., and Poole, B

Song, Y ., Sohl-Dickstein, J., Kingma, D. P., Kumar, A., Er- mon, S., and Poole, B. Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456, 2020b. Tang, Z., Hang, T., Gu, S., Chen, D., and Guo, B. Sim- plified Diffusion Schr ¨odinger Bridge.arXiv preprint arXiv:2403.14623,

arXiv 2011
[15]

Y ., Pan, J., Wu, W., et al

Wang, B., Yue, Z., Zhang, F., Chen, S., Bi, L., Zhang, J., Song, X., Chan, K. Y ., Pan, J., Wu, W., et al. Discrete visual tokens of autoregression, by diffusion, and for reasoning.arXiv preprint arXiv:2505.07538, 2025a. Wang, S., Lin, W., Huang, H., Wang, H., Cai, S., Han, W., Jin, T., Chen, J., Sun, J., Zhu, J., et al. Towards transformer-based aligned ...

arXiv
[17]

Image restoration through generalized ornstein-uhlenbeck bridge.arXiv preprint arXiv:2312.10299, 2023a

Yue, C., Peng, Z., Ma, J., Du, S., Wei, P., and Zhang, D. Image restoration through generalized ornstein-uhlenbeck bridge.arXiv preprint arXiv:2312.10299, 2023a. 11 IRBridge: Solving Image Restoration Bridge with Pre-trained Generative Diffusion Models Yue, Z., Wang, J., and Loy, C. C. Resshift: Efficient diffu- sion model for image super-resolution by re...

arXiv

Show all 23 references

[18]

Denoising dif- fusion bridge models.arXiv preprint arXiv:2309.16948,

Zhou, L., Lou, A., Khanna, S., and Ermon, S. Denoising dif- fusion bridge models.arXiv preprint arXiv:2309.16948,

arXiv
[19]

For detailed definitions of the symbols shown, please refer to the corresponding original paper

1−¯αt ( ¯αt − ¯δt)·x lq ¯βt Table 2: Reparameterization of different diffusion processes, where xlq denotes the input low-quality image. For detailed definitions of the symbols shown, please refer to the corresponding original paper. C. Critical Timsteps across Different Gener...

2020
[21]

To manage the learning rate, a constant schedule is employed with 500 warmup steps to gradually ramp up the learning rate at the start of training

To reduce GPU memory usage during training, mixed-precision training is employed. To manage the learning rate, a constant schedule is employed with 500 warmup steps to gradually ramp up the learning rate at the start of training. The scheduler includes a single cycle with a po...

2023
[22]

for dehazing. Since these datasets lack ground-truth references, we adopt widely-used no-reference image quality metrics, including MUSIQ(Ke et al., 2021), BRISQUE (Mittal et al., 2011), and NIQE(Mittal et al., 2012), to evaluate the performance of different methods. As shown ...

2021
[23]

41.975 41.870 9.587 47.586 33.879 6.881 49.888 51.574 10.054 DiffUIR(Zheng et al., 2024)41.758 43.821 9.001 49.812 32.987 5.027 50.517 50.954 8.805 DCPT(Hu et al.,

2024
[2004]

Deep retinex decomposition for low-light enhancement.arXiv preprint arXiv:1808.04560,

Wei, C., Wang, W., Yang, W., and Liu, J. Deep retinex decomposition for low-light enhancement.arXiv preprint arXiv:1808.04560,

arXiv
[2015]

Denoising diffusion implicit models.arXiv:2010.02502, October 2020a

Song, J., Meng, C., and Ermon, S. Denoising diffusion implicit models.arXiv:2010.02502, October 2020a. URL https://arxiv.org/abs/2010.02502. Song, Y . and Ermon, S. Generative modeling by estimating gradients of the data distribution.Advances in neural information processing s...

2010 arXiv
[2018]

Finding local diffusion schr\” odinger bridge using kolmogorov-arnold network.arXiv preprint arXiv:2502.19754,

Qiu, X., Yang, M., Ma, X., Li, F., Liang, D., Luo, G., Wang, W., Wang, K., and Li, S. Finding local diffusion schr\” odinger bridge using kolmogorov-arnold network.arXiv preprint arXiv:2502.19754,

arXiv
[2019]

T., Klasky, M

Chung, H., Kim, J., Mccann, M. T., Klasky, M. L., and Ye, J. C. Diffusion posterior sampling for general noisy in- verse problems.arXiv preprint arXiv:2209.14687,

arXiv
[2020]

Universal image restora- tion pre-training via degradation classification.arXiv preprint arXiv:2501.15510,

Hu, J., Jin, L., Yao, Z., and Lu, Y . Universal image restora- tion pre-training via degradation classification.arXiv preprint arXiv:2501.15510,

arXiv
[2021]

and Milanfar, P

Delbracio, M. and Milanfar, P. Inversion by direct iteration: An alternative to denoising diffusion for image restoration. arXiv preprint arXiv:2303.11435,

arXiv
[2022]

Tavt: Towards transferable audio-visual text generation

Lin, W., Jin, T., Pan, W., Li, L., Cheng, X., Wang, Y ., and Zhao, Z. Tavt: Towards transferable audio-visual text generation. InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 14983–14999, 2023a. Lin, W., Ji...

arXiv
[2023]

K., and Bovik, A

Mittal, A., Moorthy, A. K., and Bovik, A. C. Blind/referenceless image spatial quality evaluator. In 2011 conference record of the forty fifth asilomar confer- ence on signals, systems and computers (ASILOMAR), pp. 723–727. IEEE,

2011
[2024]

Bernton, E., Heng, J., Doucet, A., and Jacob, P. E. Schr ¨odinger bridge samplers.arXiv preprint arXiv:1912.13170,

1912 arXiv
[2025]

Huang, H., Xia, Y ., Ji, S., Wang, S., Wang, H., Zhu, J., Dong, Z., and Zhao, Z. Unlocking the potential of multimodal unified discrete representation through training-free code- book optimization and hierarchical alignment.arXiv preprint arXiv:2403.05168, 2024a. Huang, Z., Wu...

arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [2]

the results listed in Table 1 are obtained with fine-tuned timestep schedules, which are inherently better than the predefined schedules used in the exploratory experiments. We argue that using small-batch samples in exploratory studies is feasible because our primary focus is on observing the evolution curves of PSNR and SSIM rather than their final valu...

work page

[2] [6]

Li, B., Xue, K., Liu, B., and Lai, Y .-K

doi: 10.1109/TIP.2018.2867951. Li, B., Xue, K., Liu, B., and Lai, Y .-K. Bbdm: Image-to- image translation with brownian bridge diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern Recognition, pp. 1952–1961, 2023a. Li, T., Ku, M., Wei, C., and Chen, W. Dreamedit: Subject- driven image editing, 2023b. Li, W., Zhang, Q....

arXiv 2018

[3] [8]

Flow straight and fast: Learning to generate and transfer data with rectified flow

10 IRBridge: Solving Image Restoration Bridge with Pre-trained Generative Diffusion Models Liu, X., Gong, C., and Liu, Q. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022b. Liu, Y .-F., Jaw, D.-W., Huang, S.-C., and Hwang, J.-N. Desnownet: Context-aware deep network for snow re- mova...

arXiv

[4] [10]

Peng, X., Zheng, Z., Dai, W., Xiao, N., Li, C., Zou, J., and Xiong, H

doi: 10.1109/TPAMI.2023.3238179. Peng, X., Zheng, Z., Dai, W., Xiao, N., Li, C., Zou, J., and Xiong, H. Improving diffusion models for inverse prob- lems using optimal posterior covariance. InForty-first International Conference on Machine Learning,

arXiv 2023

[5] [12]

GenDeg: Diffusion-based Degradation Synthesis for Generalizable All-In-One Image Restoration

Rajagopalan, S., Nair, N. G., Paranjape, J. N., and Patel, V . M. Gendeg: Diffusion-based degradation synthesis for generalizable all-in-one image restoration.arXiv preprint arXiv:2411.17687,

work page Pith review arXiv

[6] [14]

P., Kumar, A., Er- mon, S., and Poole, B

Song, Y ., Sohl-Dickstein, J., Kingma, D. P., Kumar, A., Er- mon, S., and Poole, B. Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456, 2020b. Tang, Z., Hang, T., Gu, S., Chen, D., and Guo, B. Sim- plified Diffusion Schr ¨odinger Bridge.arXiv preprint arXiv:2403.14623,

arXiv 2011

[7] [15]

Y ., Pan, J., Wu, W., et al

Wang, B., Yue, Z., Zhang, F., Chen, S., Bi, L., Zhang, J., Song, X., Chan, K. Y ., Pan, J., Wu, W., et al. Discrete visual tokens of autoregression, by diffusion, and for reasoning.arXiv preprint arXiv:2505.07538, 2025a. Wang, S., Lin, W., Huang, H., Wang, H., Cai, S., Han, W., Jin, T., Chen, J., Sun, J., Zhu, J., et al. Towards transformer-based aligned ...

arXiv

[8] [17]

Image restoration through generalized ornstein-uhlenbeck bridge.arXiv preprint arXiv:2312.10299, 2023a

Yue, C., Peng, Z., Ma, J., Du, S., Wei, P., and Zhang, D. Image restoration through generalized ornstein-uhlenbeck bridge.arXiv preprint arXiv:2312.10299, 2023a. 11 IRBridge: Solving Image Restoration Bridge with Pre-trained Generative Diffusion Models Yue, Z., Wang, J., and Loy, C. C. Resshift: Efficient diffu- sion model for image super-resolution by re...

arXiv