acceptodds
Under review as a conference paper at ICLR 2027

Learning Geometry-Stabilized Intermediate Representations for Real-World Atmospheric Turbulence Restoration

Abstract

Atmospheric turbulence introduces geometric distortions and spatially varying blur, making it challenging to restore images captured over long distances.Although learning-based methods have shown promising results, their generalization to real-world scenes is often limited by reliance on synthetic training data that may not fully capture actual imaging conditions. Motivated by the geometry-stabilizing effect of temporal integration, we introduce a geometry-stabilized intermediate representation that can be learned directly from real turbulent videos while retaining residual blur. This representation enables a short-to-representation-to-clean framework that decouples single-frame restoration into geometric stabilization and detail restoration, with the latter learned from physics-guided blurred/clean pairs. We further propose a training-free algorithm that registers and refines the intermediate representations, extending the framework to a flexible number of input frames while keeping the networks fixed. Across five real-world benchmarks, our model achieves strong performance in perceptual quality, image sharpness, temporal stability, and downstream text recognition. Our code will be publicly released.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Atmospheric turbulence restoration, Long-exposure representations, Physics-guided learning, Real-world image restoration

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_y1wWOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a two-stage method for atmospheric turbulence restoration. The first stage corrects multiple input frames to estimate a stable frame. The second stage uses a Kolmogorov structure function to improve image quality. The method achieves good results on several test datasets.

Critical strengths

The experimental results are strong. The authors also provide video demo to demonstrate the method on real-world data.

Critical weaknesses

1.The novelty claims appear overstated. The proposed method has two stages. The first stage aligns multiple frames and produces a single output frame. The second stage learns a Kolmogorov structure function with adjustable parameters for turbulence restoration. The authors describe the first-stage output as a “geometry-stabilized intermediate representation.” However, this stage appears to be a network that selects a lucky frame and aligns multiple frames to it.

2.Equations (5)–(7) model the effect of exposure time on turbulence-degraded images. However, they do not appear to be connected to the model Fθ(Y)F_\theta(Y) introduced later.

3.The paper omits several important related works and datasets [1–3]. The proposed method should be compared with these works and evaluated on the relevant datasets.

4.The authors propose two metrics to evaluate temporal stability. In Appendix B.3, these metrics measure changes in image gradients between adjacent frames. However, the paper provides little explanation of how these metrics relate to the imaging model. It also lacks visual evidence to verify their reliability.

5.Since turbulence restoration is a physics-based imaging problem, the evaluation should make greater use of PSNR\SSIM, or MTF. Test charts image should also be included for qualitative visual comparisons.

6.The writing and figure legends are unclear in several places, which makes the proposed method difficult to understand.

[1] Diffeomorphic Template Registration for Atmospheric Turbulence Mitigation. CVPR24

[2] Atmospheric Turbulence Mitigation Using Complex Wavelet-Based Fusion. TIP13

[3] Single Frame Atmospheric Turbulence Mitigation: A Benchmark Study and a New Physics-Inspired Transformer Model. ECCV22

Questions for authors

1.Turbulence restoration is a physics-based imaging problem. The method design and evaluation should therefore use objective measures wherever possible. Even when reference images are unavailable and PSNR/SSIM cannot be calculated, could the authors capture images of a test chart for quantitative evaluation?

2.In some scenes, turbulence may affect only a small part of the image. How does the proposed method handle such cases, given its limited use of physical modeling?

3.Turbulence Modeling is actually a physical problem, but this paper focus on image restoration, maybe more suitable for CVPR/ICCV/ECCV?

8 comments
AuthorsOpenReview ↗

Response to Reviewer y1wW (1/4)

W1: The novelty claims appear overstated. The proposed method has two stages. The first stage aligns multiple frames and produces a single output frame. The second stage learns a Kolmogorov structure function with adjustable parameters for turbulence restoration. The authors describe the first-stage output as a “geometry-stabilized intermediate representation.” However, this stage appears to be a network that selects a lucky frame and aligns multiple frames to it.

A1: We thank the reviewer for the comments and would like to clarify the proposed framework and its main contribution.

Stage-I representation learning. Contrary to the reviewer's interpretation, Stage I neither selects a lucky frame nor aligns multiple frames at inference. Instead, it learns to predict a geometry-stabilized intermediate representation from a single turbulent frame (Eq. 1). Unlike simulation-driven approaches, Stage I learns directly from real turbulent videos without synthetic turbulence/clean-image pairs. Guided by the long-exposure analysis, temporal-average proxies and inter-frame consistency supervision enable the network to suppress geometric fluctuations while retaining residual blur for subsequent restoration (Eq. 8). Multi-frame registration is used only in the optional training-free extension (Algorithm 1).

Stage-II restoration. Our contribution is not simply the sequential application of geometric correction and deblurring, but learning an intermediate representation that connects two independently supervised restoration stages. Stage I learns geometric stabilization from real turbulent videos without paired clean images, while Stage II learns detail restoration from physics-guided synthetic blurred/clean pairs. Specifically, the Kolmogorov-inspired frequency response motivates a family of residual blur degradations, with sampled parameters and spatially varying blur (Eqs. 9–10). Stage II learns the corresponding restoration mapping rather than estimating or learning the Kolmogorov structure function itself.

Experimental support. Table 3 evaluates the learned representation using an identical, fixed Stage-II restorer. Despite requiring only a single input frame, our learned representation achieves 67.356 MUSIQ and 0.834 CPBD, compared with 58.602 MUSIQ and 0.753 CPBD when directly restoring a 50-frame temporal average. These results demonstrate that the learned representation provides a more effective restoration interface than direct temporal averaging, supporting its role in connecting real-video supervision with independently trained detail restoration.

W2: Equations (5)–(7) model the effect of exposure time on turbulence-degraded images. However, they do not appear to be connected to the model Fθ(Y) introduced later.

A2: We thank the reviewer for this question. Eqs. (5)–(7) provide the physical motivation for the intermediate representation learned by Stage I. They show that temporal integration statistically suppresses turbulence-induced geometric fluctuations while converting displacement variations into residual blur, motivating a geometry-stabilized intermediate domain.

Connection to learning. Section 3.2 translates this physical insight into the training objective of Fθ. Specifically, finite temporal averages of real static-scene turbulent videos serve as supervision proxies for the intermediate domain, while the reconstruction and inter-frame consistency losses in Eq. (8) encourage Fθ to predict geometrically stabilized representations from individual observations. Since finite-frame averages only approximate the ideal long-exposure limit, they provide scene-content anchors rather than exact physical targets. The resulting representation also connects to Stage II, where Eq. (9) extends the long-exposure frequency response in Eq. (7) to construct physics-guided blur models for independent detail-restoration training.

This connection is further supported by Table 4, where inter-frame consistency improves MUSIQ from 63.291 to 67.356, and Figure 5, which shows that the learned representations exhibit effective frequency responses compatible with the Stage-II blur family. We will clarify these connections between the physical analysis, representation learning, and restoration model in the revision.

AuthorsOpenReview ↗

Response to Reviewer y1wW (2/4)

W3: The paper omits several important related works and datasets [1–3]. The proposed method should be compared with these works and evaluated on the relevant datasets.

A3: We thank the reviewer for highlighting these relevant studies. We clarify their methodological distinctions and coverage in our evaluation.

Regarding [1], Lao et al. use optical-flow aggregation and inversion to register multiple turbulent frames, then average the aligned frames to form a template and apply blind deconvolution. In contrast, our Stage I learns directly from real turbulent videos to predict a geometry-stabilized representation from a single observation, without requiring neighboring frames or explicit registration at inference. Our key distinction lies in learning an intermediate representation that connects real-video supervision with independently trained, physics-guided detail restoration. The optional multi-frame extension further refines these predicted representations without retraining.

Regarding [2] and [3], both are already cited in our submission. The CLEAR benchmark associated with [2] is evaluated on all 13 test sequences in Table 1, although its complex-wavelet fusion method is not included as a direct baseline. TurbNet [3] is compared in Table 1, while its TurbText dataset is used for OCR evaluation in Table 2.

We acknowledge that [1] and [2] are relevant complementary approaches. Given their multi-frame input requirements and focus on static scenes, we report a matched 12-frame comparison on RLR-AT-Static for reference (Table R1). We will also expand the related-work discussion and clarify the scope of our quantitative comparisons in the revision.

Table R1. Twelve-frame comparison on RLR-AT-Static. CLEAR1 uses region mode. Lower nTII and higher scores on the other metrics indicate better performance.

MethodLIQEMUSIQNIMAMANIQAQ-AlignARNIQAHyperIQACPBDLPC-SInTIInTSS
DTR [1]1.48764.3084.7600.3792.1500.5910.5190.8510.9710.57856.373
CLEAR1 [2]2.02261.9634.7190.3482.5390.6330.5230.8080.9680.92839.697
Ours2.53767.5714.8890.4352.6840.6040.6570.8310.9760.57456.585

References

[1] Lao et al., Diffeomorphic Template Registration for Atmospheric Turbulence Mitigation, CVPR 2024.

[2] Anantrasirichai et al., Atmospheric Turbulence Mitigation Using Complex Wavelet-Based Fusion, TIP 2013.

[3] Mao et al., Single Frame Atmospheric Turbulence Mitigation: A Benchmark Study and a New Physics-Inspired Transformer Model, ECCV 2022.

W4: The authors propose two metrics to evaluate temporal stability. In Appendix B.3, these metrics measure changes in image gradients between adjacent frames. However, the paper provides little explanation of how these metrics relate to the imaging model. It also lacks visual evidence to verify their reliability.

A4: We thank the reviewer for raising this question. Appendix B.3 measures the second-order temporal differences of normalized high-pass structure maps, rather than directly measuring image gradients (Eqs. B.1–B.3). Its connection to the imaging model follows from Eq. (2): for small turbulence-induced displacements in a static scene, image variations are approximately proportional to the spatial gradient of the blurred scene and the displacement fluctuations. The second-order temporal difference therefore captures irregular structural fluctuations around scene boundaries, while normalization reduces sensitivity to high-frequency amplitude. nTII measures this temporal instability relative to the turbulent input, and nTSS is its exponential transformation. These metrics characterize temporal structural stability rather than directly measuring physical turbulence strength or distinguishing turbulence from arbitrary scene motion.

Visual and quantitative evidence. Figure 4 provides space–time traces of final restorations, while Figure 9 directly compares turbulent inputs, single-frame representations, and twelve-frame refined representations. The progressively more regular structural trajectories are consistent with Table 6, where nTII decreases from 1.000 (input) to 0.784 (Stage I) and 0.349 (twelve-frame refinement) at d=1 across 65 held-out static scenes. These results provide complementary visual and quantitative evidence that the proposed metrics reflect reductions in temporal structural fluctuations.

AuthorsOpenReview ↗

Response to Reviewer y1wW (3/4)

W5 & Q1: Since turbulence restoration is a physics-based imaging problem, the evaluation should make greater use of PSNR/SSIM, or MTF. Test charts image should also be included for qualitative visual comparisons.

A5: We thank the reviewer for this suggestion. Our primary goal is real-world atmospheric turbulence restoration, particularly under long-range imaging conditions where accurately paired turbulence-degraded and clean reference images are generally unavailable. Consequently, full-reference metrics such as PSNR and SSIM cannot be reliably applied to our main real-world benchmarks. We therefore primarily evaluate perceptual quality, structural sharpness, and temporal stability.

To provide complementary quantitative evidence, Appendix C.1 (Table 5) reports PSNR/SSIM on the paired synthetic TMT-Static dataset under an alternative supervised training protocol. Figure 5 further evaluates the effective MTFs of learned intermediate representations, demonstrating their frequency-response compatibility with the physics-guided Stage-II blur family.

Following the reviewer's suggestion, we additionally evaluate our method on two real turbulence-degraded test-chart sequences from OTIS [4], where reference images are available. As shown in Table R2, our twelve-frame method achieves the highest mean PSNR (11.025 dB) and SSIM (0.51602), together with the best temporal stability among the compared methods. We also provide two supplementary videos, Test chart1.mp4 and Test chart2.mp4, showing synchronized comparisons of the inputs, reference charts, baseline methods, our intermediate representations, and final restorations. While our method achieves better overall metrics, some baselines exhibit more discernible high-frequency details in densely patterned regions, highlighting a limitation of our training-free refinement in exploiting complementary details across frames. Future work will explore more effective cross-frame detail aggregation while preserving geometric and temporal stability.

Table R2. Additional evaluation on real OTIS test charts. Parentheses denote the number of distinct input frames.

MethodPSNR (dB) ↑SSIM ↑nTII ↓nTSS ↑
DATUM (1)10.6770.470381.07034.290
MambaTM (1)10.4710.472251.03335.602
Ours (1)10.7410.483810.83643.377
DATUM (12)10.7880.497670.62053.798
MambaTM (12)10.6980.482860.63553.013
Ours (12)11.0250.516020.55357.561

References

[4] Ferrante and Gilles, Open Turbulent Image Set (OTIS), Pattern Recognition Letters, 2017.

W6: The writing and figure legends are unclear in several places, which makes the proposed method difficult to understand.

A6: We appreciate the reviewer's feedback. We will carefully revise the manuscript to improve the clarity of the writing, figure captions, and methodological explanations, particularly the relationship between the physical analysis and the proposed framework, the distinction between single-frame inference and multi-frame refinement, and the interpretation of the evaluation metrics.

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

The paper proposes a two-stage framework for atmospheric turbulence restoration. Stage I learns a geometry-stabilized intermediate representation from real videos using temporal-average proxies and inter-frame consistency. Stage II restores details using physics-guided blurred/clean pairs. A training-free registration and refinement procedure extends the framework to multiple frames without retraining either network.

Critical strengths
  1. The framework learns stabilization without paired real clean targets while supporting single-frame inference, addressing an important obstacle in real-world turbulence restoration.

  2. Separating stabilization from detail restoration allows the two stages to use different supervision sources. Fixed-Stage-II comparisons and reuse across restoration capacities support the intermediate representation’s utility.

  3. Evaluation across real-world benchmarks and text recognition provides evidence of practical benefits. The comparisons with temporal averages and consistency-loss ablations further support the design, particularly for static scenes.

Critical weaknesses
  1. The derivation starts from a fixed clean scene (X), and Stage I is trained exclusively on RLR-AT-Static using globally translation-corrected temporal-average proxies. These assumptions do not directly cover platform motion, parallax, or changing scene content. While inference uses a single frame, the paper should explain why the learned stabilization mapping transfers to these conditions and how legitimate geometric changes are preserved.

  2. The road-sign example labeled RLR-AT-Dynamic in Figure 3 provides no temporal evidence of platform or object motion. Consecutive frames or videos showing genuine motion, together with intermediate and final outputs, are needed to assess motion preservation, geometric consistency, and possible ghosting. Temporal evaluation should distinguish turbulence suppression from suppression of legitimate scene motion.

  3. Per-dataset refinement-acceptance and bypass rates are missing. Consequently, aggregate results cannot establish how frequently the multi-frame module benefits dynamic observations versus reverting to single-frame restoration. Performance should be reported separately for accepted and bypassed cases, with matched single-frame comparisons on the same accepted frames to avoid confounding motion difficulty with refinement effectiveness.

Questions for authors
  1. Under what platform-motion and scene-change conditions is transfer from the static-scene proxy supervision expected to remain valid? What limitations remain after global translational correction?

  2. Can you provide consecutive-frame or video comparisons with clear platform or object motion, showing Stage-I representations, refined representations, and final restorations?

  3. For each dynamic dataset, please report gate acceptance/bypass rates and subgroup performance. On accepted frames, how much does refinement improve over the single-frame output evaluated on exactly the same frames?

5 comments
AuthorsOpenReview ↗

Response to Reviewer uMfY (1/3)

W1 & Q1: The derivation starts from a fixed clean scene (X), and Stage I is trained exclusively on RLR-AT-Static using globally translation-corrected temporal-average proxies. These assumptions do not directly cover platform motion, parallax, or changing scene content. While inference uses a single frame, the paper should explain why the learned stabilization mapping transfers to these conditions and how legitimate geometric changes are preserved. Under what platform-motion and scene-change conditions is transfer from the static-scene proxy supervision expected to remain valid? What limitations remain after global translational correction?

A1: We thank the reviewer for raising this important question regarding transfer from static-scene supervision to dynamic observations.

Transfer rationale. The static-scene assumption in Secs. 3.1–3.2 enables temporal-average supervision from real turbulent videos without paired clean images, while global translational correction reduces camera drift during proxy construction. Importantly, Stage I processes each frame independently (Eq. 1), without aligning observations to a fixed scene reference. For sufficiently short exposures, turbulence-induced distortions and blur still act on the current scene configuration, even when objects or camera viewpoints change over time. This provides a rationale for applying the learned stabilization mapping to dynamic observations, with the aim of suppressing turbulence-induced distortions while preserving legitimate scene geometry.

Empirical evidence. Despite being trained exclusively on RLR-AT-Static, our single-frame model achieves strong performance on RLR-AT-Dynamic (Table 1), obtaining 66.743 MUSIQ and 0.867 nTII, compared with 65.761 MUSIQ for TurbNet and 0.953 nTII for AT-Net, the strongest respective baselines. To further examine spatial consistency, we evaluate the frozen Stage-I model under controlled global translations by comparing F(T(Y)) with T(F(Y)), where T denotes an imposed integer-pixel translation. Table R1 reports consistently high agreement (SSIM ≥ 0.9999) at translations of 4, 8, 16, and 32 pixels, demonstrating approximate translation equivariance rather than alignment to a fixed spatial reference. Complementary qualitative evidence is provided by the consecutive-frame examples involving moving trains and road vehicles in our supplementary video (video results.mp4), as further discussed in W2.

Applicability and limitations. We further investigate sensitivity to temporal averaging on RLR-AT-Dynamic (Table R2). With the restoration networks unchanged, MUSIQ decreases modestly from 66.743 at N=1 to 66.016 at N=4, but drops substantially to 52.680 at N=36. This suggests that prolonged temporal averaging of dynamic observations can introduce motion-related blur and scene-content mixing that degrade restoration performance. More generally, substantial within-exposure motion blur, depth-dependent parallax, and complex occlusions remain challenging, potentially compromising geometric stabilization and detail restoration in dynamic scenes. Regarding training supervision, global translational correction is applied only during proxy construction from static-scene videos. Although it reduces camera drift, residual non-translational misalignment may still affect proxy quality.

We will clarify the transfer rationale, supporting evidence, and the corresponding limitations in the revision.

Table R1. Stage-I consistency under controlled global translations. Compare F(T(Y)) with T(F(Y)) on 65 held-out scenes, using five frames per scene and four translation directions (1,300 pairs per row). Metrics use the common central 384 × 384 region.

Translation magnitude (pixels)PSNR (dB) ↑SSIM ↑
481.8090.999998
878.3590.999996
1674.3780.999990
3269.5840.999972

Table R2. Effect of temporal averaging on RLR-AT-Dynamic (N = 1, 2, 4, 36). Average N consecutive input frames, then apply Stage I and Stage II. Full-image evaluation uses the same 60 scenes and 6,000 outputs per setting.

NLIQE ↑MUSIQ ↑NIMA ↑MANIQA ↑Q-Align ↑ARNIQA ↑HyperIQA ↑
12.12766.7434.7220.3692.7710.6050.631
22.05766.3014.7260.3642.7450.6020.628
42.01366.0164.7480.3612.7060.6010.625
361.63652.6804.2080.2992.2710.5690.497
AuthorsOpenReview ↗

Response to Reviewer uMfY (2/3)

W2 & Q2: The road-sign example labeled RLR-AT-Dynamic in Figure 3 provides no temporal evidence of platform or object motion. Consecutive frames or videos showing genuine motion, together with intermediate and final outputs, are needed to assess motion preservation, geometric consistency, and possible ghosting. Temporal evaluation should distinguish turbulence suppression from suppression of legitimate scene motion. Can you provide consecutive-frame or video comparisons with clear platform or object motion, showing Stage-I representations, refined representations, and final restorations?

A2: We thank the reviewer for this valuable suggestion. We agree that the single-frame comparison in Fig. 3 alone cannot demonstrate temporal motion preservation. We therefore first clarify the scope of our framework and its motion-handling strategy, followed by consecutive-frame evidence of turbulence suppression and legitimate motion preservation.

Method scope and motion handling. Our primary focus is single-frame atmospheric turbulence restoration, where each observation is processed independently without temporal alignment. We further explore a training-free multi-frame extension to reduce residual geometric distortions and improve temporal stability when additional observations are available. Specifically, Algorithm 1 employs a RANSAC-based motion gate to bypass temporal refinement under local translations or rotations. For accepted windows, registration and bounded local correction further stabilize intermediate representations, while reference-guided refinement corrects low-frequency structures and retains high-frequency details from the target frame (Eq. 13). This strategy aims to improve temporal stability while limiting unintended changes to dynamic scene content.

Real-motion evidence. Our supplementary video (video results.mp4) provides consecutive-frame comparisons involving genuine object and camera motion. In Case 2 (00:21–00:38), the single-frame examples show a moving train and road vehicles, whose changing positions are retained in the Stage-I representations and final restorations while turbulence-induced fluctuations along object boundaries are reduced, without obvious persistent ghosting. In Case 4 (00:55–01:02), the twelve-frame example shows coherent displacement of foreground terrain, industrial buildings, and distant hills due to camera motion. The intermediate representations and final restorations retain these viewpoint changes while producing more regular structural boundaries. To quantitatively support these observations, we additionally provide new trajectory visualizations for the moving-object sequences in Case 2, showing the tracked centers, trajectories, and contours of a train window and a tanker in the two scenes, respectively (see “case2_train_window_trajectory.png” and “case2_tanker_trajectory.png” in the updated supplementary material). Each figure compares the input, Stage-I representation, and final output at matched time points. From the tracked centers, we measured displacement from the first to the last frame and computed the cosine similarity between the full input and final trajectories. From the contours, we measured the mean RMS difference between adjacent-frame contour residuals. As shown in Table R3, the final trajectories closely follow the input trajectories in both scenes, with displacement differences below 0.4% and cosine similarities above 0.9999. Meanwhile, the final contour residual RMS drops by about 78% for the train window and 7% for the tanker, with most of the reduction already present in the Stage-I representations. For the camera-motion sequence in Case 4, the estimated global-motion trends remain closely aligned between the input and the restored outputs, with a trajectory cosine similarity of 0.99981. The motion gate accepts all evaluated target frames. We further compare the Stage-I and refined representations at the same frame indices and in the same coordinate system, observing a 16.1% decrease in nTII, from 0.813 to 0.682. The video comparison is provided in the updated supplementary material (“case4_single_vs_12_frames.mp4”). Together, the qualitative comparisons and quantitative measurements provide complementary evidence of turbulence suppression while preserving observable object and camera motion.

Table R3. Motion preservation and contour fluctuations in Case 2. The displacement ratio is Final / Input. Contour residual RMS is reported in the order Input → Stage I → Final.

Clip / targetFramesInput disp. (px)Final disp. (px)Trajectory cosineDisp. ratioContour RMS (px)
Train / window9122.412122.3340.9999999.94%0.579 → 0.113 → 0.127
Vehicles / tanker25109.308108.8830.9999599.61%0.700 → 0.648 → 0.651
AuthorsOpenReview ↗

Response to Reviewer uMfY (3/3)

W3 & Q3: Per-dataset refinement-acceptance and bypass rates are missing. Consequently, aggregate results cannot establish how frequently the multi-frame module benefits dynamic observations versus reverting to single-frame restoration. Performance should be reported separately for accepted and bypassed cases, with matched single-frame comparisons on the same accepted frames to avoid confounding motion difficulty with refinement effectiveness. For each dynamic dataset, please report gate acceptance/bypass rates and subgroup performance. On accepted frames, how much does refinement improve over the single-frame output evaluated on exactly the same frames?

A3: We thank the reviewer for this suggestion. To distinguish the applicability of temporal refinement from its effectiveness, we further evaluate the motion-gate acceptance/bypass rates and compare single-frame and twelve-frame restoration on matched observations across RLR-AT-Dynamic, UoB-Turb, and CLEAR (Table R4).

For the reported accepted groups, temporal refinement substantially improves stability over single-frame restoration, reducing nTII by 55.3%, 50.9%, and 34.6% on the three datasets, respectively, with corresponding nTSS improvements of 64.9%, 55.3%, and 34.1%. Bypassed observations follow the original single-frame restoration pathway and receive no temporal refinement.

These results demonstrate the effectiveness of temporal refinement when the motion gate permits its application, while also highlighting that its practical coverage depends on scene motion. However, as discussed in the manuscript's Limitations paragraph, the current extension uses training-free registration and reference-guided refinement of the intermediate representations primarily to improve temporal stability, with all seven perceptual scores changing by less than 5% relative to their single-frame values in each Accept group (Table R4). Exploring learned temporal aggregation within the geometry-stabilized domain to better exploit cross-frame dependencies and complementary details, particularly for frames that currently bypass refinement, is an important direction for future work. We will include the subgroup evaluation protocol and clarify the gate's limitations in the revision.

Table R4. Restoration quality and temporal stability by scene-level evaluation group. Accept and bypass denote the assigned scene-level comparison groups. Rates give their proportions within each dataset. Single-frame and twelve-frame scores use identical target frames. Temporal metrics use d=1. Entries show single-frame → twelve-frame (relative change from single-frame).

\scriptsize \begin{array}{ccccccccccc}\text{Dataset}&\text{Group / Rate}&\text{LIQE ↑}&\text{MUSIQ ↑}&\text{NIMA ↑}&\text{MANIQA ↑}&\text{Q-Align ↑}&\text{ARNIQA ↑}&\text{HyperIQA ↑}&\text{nTII ↓}&\text{nTSS ↑}\\\\\hline\text{RLR-AT-Dynamic}&\text{Bypass}&2.374&66.879&4.829&0.382&2.830&0.614&0.639&0.854&42.688\\\\&66.67\\%&→&→&→&→&→&→&→&→&→\\\\&&2.374&66.879&4.829&0.382&2.830&0.614&0.639&0.854&42.688\\\\&&(0.0\\%)&(0.0\\%)&(0.0\\%)&(0.0\\%)&(0.0\\%)&(0.0\\%)&(0.0\\%)&(0.0\\%)&(0.0\\%)\\\\\hline\text{RLR-AT-Dynamic}&\text{Accept}&1.635&66.471&4.509&0.345&2.653&0.586&0.617&0.895&40.932\\\\&33.33\\%&→&→&→&→&→&→&→&→&→\\\\&&1.566&66.995&4.531&0.354&2.642&0.587&0.622&0.400&67.517\\\\&&(-4.2\\%)&(+0.8\\%)&(+0.5\\%)&(+2.6\\%)&(-0.4\\%)&(+0.2\\%)&(+0.8\\%)&(-55.3\\%)&(+64.9\\%)\\\\\hline\text{UoB-Turb}&\text{Bypass}&2.199&53.137&4.739&0.322&3.006&0.574&0.443&0.876&41.673\\\\&15.79\\%&→&→&→&→&→&→&→&→&→\\\\&&2.199&53.137&4.739&0.322&3.006&0.574&0.443&0.876&41.673\\\\&&(0.0\\%)&(0.0\\%)&(0.0\\%)&(0.0\\%)&(0.0\\%)&(0.0\\%)&(0.0\\%)&(0.0\\%)&(0.0\\%)\\\\\hline\text{UoB-Turb}&\text{Accept}&2.385&56.035&4.614&0.325&2.892&0.588&0.471&0.863&42.506\\\\&84.21\\%&→&→&→&→&→&→&→&→&→\\\\&&2.375&55.702&4.685&0.336&2.956&0.596&0.475&0.423&65.994\\\\&&(-0.4\\%)&(-0.6\\%)&(+1.5\\%)&(+3.2\\%)&(+2.2\\%)&(+1.4\\%)&(+1.0\\%)&(-50.9\\%)&(+55.3\\%)\\\\\hline\text{CLEAR}&\text{Bypass}&1.758&57.655&4.568&0.240&3.037&0.553&0.477&0.876&41.994\\\\&30.77\\%&→&→&→&→&→&→&→&→&→\\\\&&1.758&57.655&4.568&0.240&3.037&0.553&0.477&0.876&41.994\\\\&&(0.0\\%)&(0.0\\%)&(0.0\\%)&(0.0\\%)&(0.0\\%)&(0.0\\%)&(0.0\\%)&(0.0\\%)&(0.0\\%)\\\\\hline\text{CLEAR}&\text{Accept}&1.632&44.452&4.710&0.243&2.120&0.498&0.379&0.855&43.157\\\\&69.23\\%&→&→&→&→&→&→&→&→&→\\\\&&1.697&44.045&4.704&0.247&2.124&0.509&0.381&0.559&57.868\\\\&&(+4.0\\%)&(-0.9\\%)&(-0.1\\%)&(+1.6\\%)&(+0.2\\%)&(+2.2\\%)&(+0.5\\%)&(-34.6\\%)&(+34.1\\%)\end{array}

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

To address image degradation in complex atmospheric turbulence environments, this paper designs a phased image restoration method, with the core idea being as follows decoupling geometric distortion mitigation from detail recovery by learning a geometrically stabilized intermediate representation. The first stage is trained on real static turbulence videos, learning a mapping from single-frame turbulence observations to a geometrically more stable intermediate representation. The second stage learns to recover residual details from the intermediate representation via a generalized frequency attenuation model with spatially varying blur severity. Experiments cover five real-world benchmarks, reporting strong non-reference perceptual quality, sharpness, and temporal stability results.

Critical strengths
  1. The first stage constructs supervision surrogates from real turbulence videos without relying on strictly paired real clean and turbulent images. This idea effectively alleviates the difficulty of obtaining real paired data and is very valuable.
  2. The multi-frame extension is training-free, offering practical utility.
Critical weaknesses
  1. The short-to-representation-to-clean framework proposed by the authors is highly similar in concept to the Detilt-then-Deblur approach proposed by Fan et al.[1]. Therefore, this general decomposition itself lacks sufficient novelty as the primary original contribution of the current manuscript.
  2. Stage 1 and Stage 2 adopt different dataset simulation strategies, how are Stage 1 and Stage 2 bridged? In other words, how is it guaranteed that the outputs of Stage 1 match the distribution of the simulated training data in Stage 2? How do the authors prove that the local degradations, geometric residuals, noise, and texture statistics of real Stage I outputs are well-covered by the training distribution of Stage 2?
  3. Why does the proposed method underperform TMT across all three OCR recognizers on TurbText? Does this stem from over-smoothing in Stage I or erroneous detail hallucinations in Stage 2?
  4. The multi-frame extension is training-free. If occlusions or disappearing objects occur, how is inter-frame consistency ensured? If multi-frame averaging of the same static scene is used as a surrogate target, is the cross-frame consistency loss still effective? Furthermore, is this method still applicable to rapidly changing dynamic scenes?

[1] Degradation & Restoration: A Low-cost Pipeline for Long-range Single-frame Turbulence Mitigation https://openreview.net/forum?id=6pHsT7rLJN

Questions for authors

Please refer to the Weaknesses.

8 comments
AuthorsOpenReview ↗

Response to Reviewer mvSp (1/3)

W1: The short-to-representation-to-clean framework proposed by the authors is highly similar in concept to the Detilt-then-Deblur approach proposed by Fan et al.[1]. Therefore, this general decomposition itself lacks sufficient novelty as the primary original contribution of the current manuscript.

A1: We thank the reviewer for highlighting the relevant work by Fan et al. [1]. We agree that separating geometric correction from deblurring has precedents and do not claim novelty in the two-stage architecture itself. Our contribution lies in learning a geometry-stabilized intermediate representation from real turbulent videos to enable representation-level decomposition under heterogeneous supervision.

Unlike [1], which relies on synthetic turbulence supervision, our Stage I learns a single-frame geometry-stabilized representation directly from real turbulent videos using temporal-average proxies and inter-frame consistency (Sec. 3.2, Eq. 8), without synthetic geometric supervision or paired real clean images. This learned representation bridges heterogeneous supervision sources, enabling real-video-supervised geometric stabilization and independently trained physics-guided detail restoration.

The effectiveness of real-video supervision is supported by Table 4 (real vs. synthetic P2S supervision: 62.277 vs. 57.004 MUSIQ under matched losses). Table 3 further demonstrates the utility of the learned representation (learned representation vs. 50-frame average: 67.356 vs. 58.602 MUSIQ with fixed Stage II).

We will clarify this distinction and explicitly discuss Fan et al. in the revision.

[1] Fan et al. Degradation & Restoration: A Low-cost Pipeline for Long-range Single-frame Turbulence Mitigation. ICLR 2026 submission.

W2: Stage 1 and Stage 2 adopt different dataset simulation strategies, how are Stage 1 and Stage 2 bridged? In other words, how is it guaranteed that the outputs of Stage 1 match the distribution of the simulated training data in Stage 2? How do the authors prove that the local degradations, geometric residuals, noise, and texture statistics of real Stage I outputs are well-covered by the training distribution of Stage 2?

A2: We thank the reviewer for raising this important question about cross-stage compatibility. While exact distribution matching between real Stage-I outputs and synthetic Stage-II inputs is not guaranteed, the two stages are connected through a geometry-stabilized intermediate representation. Specifically, Stage I uses temporal-average proxies and inter-frame consistency to suppress observation-dependent geometric variations while retaining residual blur (Sec. 3.2). Stage II is independently trained on a diversified blur family with spatially varying strengths and frequency-decay profiles spanning Kolmogorov and Gaussian cases (Sec. 3.3), designed to accommodate variations in the residual blur of learned representations.

We provide complementary empirical evidence supporting this cross-stage compatibility. Figure 5 shows substantial overlap between the effective edge-based MTFs of real Stage-I outputs and the synthetic blur family, supporting frequency-response compatibility. With Stage II fixed, Table 3 demonstrates that learned single-frame representations outperform 50-frame temporal averages (67.356 vs. 58.602 MUSIQ), supporting their effectiveness as inputs to the independently trained restorer. Table 4 further shows that spatially varying blur and diverse frequency-decay profiles improve restoration performance on real data, supporting the Stage-II design.

Together, these results support practical compatibility between the independently trained stages, rather than complete distribution coverage of geometric residuals, sensor noise, and texture statistics. We will clarify this distinction in the revision.

AuthorsOpenReview ↗

Response to Reviewer mvSp (2/3)

W3: Why does the proposed method underperform TMT across all three OCR recognizers on TurbText? Does this stem from over-smoothing in Stage I or erroneous detail hallucinations in Stage 2?

A3: We thank the reviewer for raising this question. As discussed in Sec. 4.2 and Fig. 4, severe turbulence may cause Stage I to produce geometrically stabilized but blurrier representations, weakening fine character details. To investigate this, we conduct a stage-wise OCR analysis on 500 matched word regions from 100 TurbText scenes.

For CRNN/DAN/ASTER, recognition accuracy changes from 23.8/34.8/38.8% on the original inputs to 23.2/28.2/31.0% after Stage I, suggesting that geometric stabilization alone may not sufficiently preserve fine character cues. Accuracy subsequently improves to 68.4/71.6/77.2% after the full restoration pipeline. Among 116/141/155 words correctly recognized after Stage I, only 0/1/0 become incorrect in the final outputs, suggesting that downstream restoration rarely corrupts already recognizable text.

The remaining gap relative to TMT may also reflect differences in temporal information utilization. TMT employs temporal-channel attention to aggregate complementary features across frames, whereas our multi-frame refinement primarily corrects residual geometry, with detail restoration performed on a single refined representation. Among words recognized by TMT but not by ours, 85.9/77.8/77.8% are recognizable in at most 20% of the original frames, consistent with weak character cues and the potential importance of cross-frame detail aggregation.

These findings suggest that insufficient preservation of fine character details in Stage I, together with limited exploitation of complementary temporal information, may contribute to the remaining OCR gap. We will incorporate this analysis and clarify the corresponding limitations in the revision.

AuthorsOpenReview ↗

Response to Reviewer mvSp (3/3)

W4: The multi-frame extension is training-free. If occlusions or disappearing objects occur, how is inter-frame consistency ensured? If multi-frame averaging of the same static scene is used as a surrogate target, is the cross-frame consistency loss still effective? Furthermore, is this method still applicable to rapidly changing dynamic scenes?

A4: We thank the reviewer for raising these important questions. Our primary focus is single-frame atmospheric turbulence restoration. We further explore a training-free multi-frame extension to reduce residual geometric distortions and improve temporal stability when additional observations are available, without introducing a learned temporal aggregation module or retraining either network.

Regarding occlusions and scene changes, Algorithm 1 employs a RANSAC-based motion gate that bypasses temporal refinement under local translations or rotations. For accepted windows, global registration and bounded local correction further stabilize the intermediate representations. Importantly, the refinement uses a temporal reference to correct low-frequency structures while retaining high-frequency details from the target frame (Eq. 13), limiting reliance on neighboring frames for fine-detail reconstruction. Although the current extension does not explicitly model occlusions or local object motion, the independent single-frame restoration pathway remains available without requiring temporal correspondence.

Regarding inter-frame consistency, Stage I uses shared temporal-average proxies for static-scene training. While perfect reconstruction of the shared proxy would make the consistency loss redundant, finite training does not guarantee identical predictions from different turbulent observations. The reconstruction loss anchors predictions to the shared proxy, whereas the consistency loss directly suppresses observation-dependent variations (Sec. 3.2, Eq. 8). To further validate its effectiveness, we extend the ablation in Table 4 to twelve-frame inference (Table R1). We evaluate temporal stability using nTII, which measures normalized second-order temporal fluctuations of image structures, and its corresponding stability score nTSS. Under the same fixed Stage II, consistency improves MUSIQ from 63.291 to 67.356, reduces nTII from 0.952 to 0.838, and increases nTSS from 38.930 to 43.646 in single-frame inference. In twelve-frame inference, MUSIQ improves from 65.600 to 67.571, nTII decreases from 0.634 to 0.574, and nTSS increases from 53.321 to 56.585. These within-setting comparisons demonstrate that consistency regularization remains effective despite shared proxy supervision and complements inference-time refinement.

Table R1. Effect of inter-frame consistency under single-frame and twelve-frame inference. Evaluated on 65 RLR-AT-Static scenes (6,500 frames per setting), using the same fixed Stage II.

Input framesConsistencyLIQE ↑MUSIQ ↑NIMA ↑MANIQA ↑Q-Align ↑ARNIQA ↑HyperIQA ↑nTII ↓nTSS ↑
1✗2.17963.2914.6200.3792.6230.5830.5850.95238.930
1✓2.53967.3564.8590.4272.6720.6040.6560.83843.646
12✗2.35165.6004.7460.4052.6770.5910.6190.63453.321
12✓2.53767.5714.8890.4352.6840.6040.6570.57456.585

Regarding rapidly changing dynamic scenes, our single-frame pipeline processes each observation independently without requiring temporal correspondence. The multi-frame extension is primarily intended to further stabilize persistent scene structures when reliable registration is possible. The supplementary video (video results.mp4) provides qualitative examples of single-frame restoration in dynamic scenes (Case 2, 00:21–00:38) and twelve-frame restoration with camera motion (Case 4, 00:55–01:02). Robust handling of complex occlusions and rapid visibility changes remains challenging for the current training-free refinement and is an important direction for future work.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Final Remarks by Authors

Dear Reviewers,

As the Author–Reviewer discussion period draws to a close, we sincerely thank all three reviewers for their thoughtful feedback, constructive suggestions, and time devoted to evaluating our work.

We greatly appreciate the reviewers' recognition of the strengths of our submission. Reviewer mvSp recognized the value of learning from real turbulent videos without paired clean targets and the practical utility of the training-free multi-frame extension. Reviewer uMfY highlighted the intermediate representation's role in bridging heterogeneous supervision sources, its effectiveness supported by fixed-Stage-II comparisons, and the practical benefits demonstrated across real-world benchmarks. Reviewer y1wW acknowledged the strong experimental results and real-world video demonstrations. We sincerely appreciate these encouraging assessments and the constructive suggestions that have helped us further strengthen our work.

We would like to reiterate the three central contributions of our work:

  1. Learning geometry-stabilized representations from real videos. Stage I learns to predict an intermediate image from a single turbulent observation using temporal-average proxies and inter-frame consistency, without requiring paired real clean targets.

  2. Representation-level decomposition under heterogeneous supervision. Motivated by long-exposure analysis, the learned intermediate representation connects geometric stabilization learned from real turbulent videos with independently trained, physics-guided detail restoration.

  3. Training-free multi-frame extension. Registration and refinement of the learned representations further improve temporal stability when additional observations and reliable alignment are available, without retraining either restoration network.

Throughout the rebuttal, we have carefully considered the reviewers' concerns and provided additional experiments, analyses, and methodological clarifications, including:

  1. Intermediate representation and supervision design. We clarified the image-domain nature of the Stage-I output, its physical motivation, and the rationale for temporal-average supervision. We also distinguished the ideal long-exposure image from its finite-frame supervision proxy and explained the complementary roles of reconstruction and inter-frame consistency, supported by the comparisons and ablations in Tables 3 and 4.

  2. Cross-stage compatibility and limitations. We conducted a controlled comparison with a fixed Stage-I network and two Stage-II training strategies. The results show that training on matched Stage-I outputs further improves restoration performance, providing additional insight into cross-stage compatibility while identifying opportunities to better model residual degradations and investigate their impact on real-world restoration fidelity.

  3. Motion preservation and temporal refinement. We clarified the distinction between single-frame restoration and optional training-free multi-frame refinement. Additional videos, trajectory analyses, and quantitative measurements provide evidence of object and camera motion preservation. We also reported refinement acceptance/bypass statistics and matched subgroup comparisons to assess both the effectiveness and applicability of temporal refinement.

  4. Broader experimental validation. We added full-reference PSNR/SSIM evaluations and synchronized video comparisons on real OTIS test charts, comparisons with additional registration- and fusion-based methods, and an analysis of spatially varying turbulence. We also investigated text-restoration limitations through stage-wise OCR analysis and provided further visual evidence in the updated supplementary material.

We recognize that cross-stage distribution mismatch and generalization under more challenging real-world conditions remain important directions for further investigation. We believe our work offers a distinct perspective on predominantly simulation-driven turbulence restoration by integrating real-video representation learning with physics-guided restoration, achieving strong performance across real-world benchmarks (RLR-AT-Static, RLR-AT-Dynamic, UoB-Turb, CLEAR, and TurbText), with additional validation on real OTIS test charts. More broadly, this perspective may inspire further exploration of real-data-driven learning and heterogeneous supervision for atmospheric turbulence restoration.

We sincerely hope that our responses and supplementary evidence have helped clarify the motivation, contributions, and practical value of our work. We would be grateful for your consideration of the additional evidence and clarifications provided during the rebuttal as part of the overall evaluation of our work.

Thank you again for your time, constructive feedback, and careful consideration.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.