acceptodds
Under review as a conference paper at ICLR 2027

When Motion Misleads: Causal-Role-Aware Motion-Turbulence Disentanglement for Turbulent Infrared Video Super-Resolution

Abstract

Infrared turbulent video super-resolution is fundamentally challenging due to the entanglement of real object/camera motion and random turbulence-induced distortions. Existing turbulence mitigation and video super-resolution methods often conflate turbulence-induced pseudo-motion with real object/camera motion and model them as a unified temporal correspondence signal, leading to spurious alignments and unstable details. We propose CAMO, a Causal-role-Aware Motion-turbulence disentanglement framework for One-step diffusion restoration, rooted in the physical principle that the real motion induces coherent phase transport in the infrared domain, while turbulence-induced refractive perturbations break such coherence. This coherence provides a reliability cue for factorizing mixed temporal dynamics into constructive object motion and destructive turbulence components. The turbulence component is used to suppress local non-rigid distortions, while the real motion component promotes temporal consistency and enhances motion-aware details during reconstruction. Furthermore, CAMO regulates the generative conditioning and velocity prediction based on motion reliability, enabling constructive motion cues to be selectively exploited while preventing the amplification of uncertain turbulence-induced pseudo-motion. This unified design enables end-to-end turbulence suppression, temporal information utilization, and high-resolution reconstruction. Extensive experiments under complex infrared turbulence demonstrate improved restoration fidelity, temporal stability, and robustness over existing methods. Anonymous code is available at https://camo23333.github.io/camo.github.io/CAMO.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Infrared video super-resolution, Infrared imaging, Low-level vision

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_sqqUOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies infrared video super-resolution (IVSR) under turbulent conditions. The authors argue that real object/camera motion can provide effective cross-frame information, whereas the pseudo-motion produced by atmospheric turbulence instead disrupts alignment, and therefore the two should not be treated uniformly as ordinary motion. CAMO uses LMTD in the latent space to factorize the temporal variations into an object-motion component and a turbulence component, and designs TRMCP to estimate motion reliability based on phase, radiative, and edge consistency. Subsequently, TCLR is used to suppress turbulence perturbations, MCTE exploits the real motion to enhance temporal information, and MRGD adjusts the DiT features and the prediction process according to reliability. The model is built on CogVideoX1.5 and adopts one-step inference.

Critical strengths
  • The problem definition is more targeted than conventional motion-guided VSR. The paper does not simply pursue a more accurate optical flow; instead, it distinguishes between "beneficial real motion" and "harmful turbulence-induced pseudo-motion." For turbulent IVSR, this entry point is reasonable and does indeed target the tendency of existing methods to misuse temporal correspondence.
  • The overall method is fairly complete, with clear division of labor among modules. LMTD handles the factorization, TCLR processes the turbulence component, MCTE uses the object-motion component, and MRGD further adjusts the generation process according to reliability. Although there are many modules, they are not simply stacked: each one basically corresponds to a problem raised earlier.
  • The results on FLIR-IVSR are strong. On this dataset, CAMO achieves the best results on PSNR, SSIM, LPIPS, DISTS, MUSIQ, BRISQUE, DOVER, and FasterVQA, and it does not merely dominate on perceptual metrics—reconstruction metrics and temporal metrics are also strong simultaneously. The advantage on M3FD is not as comprehensive, but the results remain competitive overall.
Critical weaknesses
  • The claim of "causal-role-aware" is somewhat overstated. In fact, the paper does not perform causal identification in the strict sense. M=Mobj+MturbM = M_{\mathrm{obj}} + M_{\mathrm{turb}} is factorized by learnable modules and then constrained through HQ/LQ consistency, TRMCP, and reconstruction losses. Such a design can produce useful representations, but it does not guarantee that the two components respectively correspond to object motion and turbulence in the physically meaningful sense.
  • The basic assumption of TRMCP may not be robust under complex motion. The method first estimates local translation and then treats phase, radiative, and edge residuals that cannot be explained by coherent translation as turbulence evidence. However, fast non-rigid motion, occlusion, rotation, and scale changes can also produce large residuals; conversely, slowly varying turbulence does not necessarily manifest as obvious random high-frequency perturbations. Thus, the current prior resembles an empirically effective reliability cue rather than a physical criterion sufficient to distinguish the two types of motion.
  • The experiments on M3FD cannot actually sufficiently validate motion-turbulence disentanglement. The authors repeat a single image for 30 frames and then add turbulence, downsampling, and noise. The videos generated in this way have no real object motion themselves, so this dataset mainly validates turbulence suppression, rather than whether the two can be correctly separated when "real motion and pseudo-motion coexist."
  • Backbone capacity is an obvious confounding factor. CAMO has about 5.8B parameters, far more than HATIR, DLoRAL, DGAF-VSR and other methods, and is built directly on CogVideoX1.5. The current results can show that the final system is strong, but it is not yet easy to determine how much of the gain comes from motion-turbulence modeling and how much comes from the larger pretrained video backbone.
Questions for authors

The M3FD videos are generated by repeating a static image, so they contain no real motion. Could the authors add a test set that contains both real scene motion and real/simulated atmospheric turbulence? This is important for validating the paper's core disentanglement claim.

1 comment
AuthorsOpenReview ↗

Response to Reviewer sqqU

We thank the reviewer for the thoughtful assessment and for recognizing the motivation for treating object motion and turbulence-induced pseudo-motion differently and the distinct roles of the proposed modules. We address each point below.

Q1. Meaning of “causal-role-aware” and interpretation of the components. We thank the reviewer for drawing attention to the scope of this term. In CAMO, “causal-role-aware” refers to the distinct roles of the learned components in restoration. The turbulence component guides LR latent rectification before diffusion, while the object-motion component guides cross-frame refinement after diffusion. The training objectives encourage this separation without imposing a unique physical decomposition. Fig. 11 (Appendix A.1) shows that the object-motion component tends to follow coherent structures and moving boundaries, while the turbulence component exhibits more spatially irregular responses. In Figs. 10 and 19, injecting the learned turbulence component into HR videos produces SynLR sequences that more closely match the observed LR sequences than either the original HR videos or the random blur-and-distortion control. These results support the interpretation that the learned component captures turbulence-related degradation patterns.

Q2. Reliability of TRMCP under complex motion. We agree that occlusions, rotations, scale changes, and fast nonrigid motion can leave large residuals after translation compensation, while slowly varying turbulence may elicit weaker responses. TRMCP estimates translation separately in overlapping patches so that compensation can adapt to spatially varying motion. The resulting phase, radiative, and edge residuals measure local coherence, and CAMO combines their fused score with latent motion uncertainty to estimate reliability (Sec. 3.2). Fig. 12 (Appendix A.3) and our additional analysis in Table R1 show that the mean fused score increases with turbulence strength over the tested range.

Table R1. TRMCP component responses across turbulence strengths. Values are mean ± sample standard deviation across 50 videos at each setting. Normalization scales are held fixed across settings. Smaller r0r_0 indicates stronger turbulence.

r0r_0Phase residualRadiative residualEdge residualQtherQ_{\mathrm{ther}}
0.0250.10881 ± 0.022360.08192 ± 0.019910.06334 ± 0.009870.08469 ± 0.01398
0.0350.09504 ± 0.022080.06936 ± 0.015820.05827 ± 0.008360.07422 ± 0.01169
0.0500.08071 ± 0.021440.05994 ± 0.014630.05426 ± 0.008120.06497 ± 0.01049
0.0710.06914 ± 0.020180.05251 ± 0.015440.05005 ± 0.009070.05724 ± 0.01001
0.1000.06220 ± 0.018490.04738 ± 0.019720.04565 ± 0.011570.05174 ± 0.01078
0.1400.05655 ± 0.016460.04273 ± 0.020050.04134 ± 0.012970.04687 ± 0.01143

Q3. Evaluation when real motion and turbulence coexist. We agree that the repeated-frame M³FD sequences evaluate restoration under turbulence in static scenes. FLIR-IVSR includes captured infrared sequences with real scene motion (camera and object motion) and paired turbulent LR inputs. Table 1 reports restoration performance across the test split, and Figs. 5 and 16 show temporal profiles for individual FLIR-IVSR examples. The ablation in Fig. 7 further measures the restoration benefit of the decomposition across the same split. The video demos on our anonymous project page provide additional qualitative examples of CAMO preserving moving object boundaries and reducing temporal jitter in FLIR-IVSR clips with real object and camera motion under turbulence.

Q4. Influence of the pretrained backbone. We thank the reviewer for raising the question of how much of the improvement is attributable to the CAMO design. We added a one-step restoration baseline using the same pretrained CogVideoX1.5 backbone (Table 3, Appendix A.10). Full CAMO improves PSNR from 31.6642 to 36.7342 dB (+5.07 dB) and reduces LPIPS from 0.1642 to 0.1160, with improvements in SSIM and BRISQUE as well. This comparison measures the gain of the full CAMO design with the pretrained backbone held fixed. The LMTD ablation in Fig. 7 and the separate TCLR and MCTE ablations in Table 3 provide additional evidence about the contributions of these components within the pipeline.

ConfigurationPSNR ↑SSIM ↑LPIPS ↓BRISQUE ↓
CogVideoX1.5 baseline31.66420.89560.164251.1095
CAMO w/o TCLR and MCTE35.11800.93180.131044.8297
CAMO w/o TCLR35.96700.93200.121342.2750
CAMO w/o MCTE36.09360.94020.124441.5441
Full CAMO36.73420.94770.116038.0304
Reviewer_TtVyOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a causal-role-aware framework for turbulent infrared video super-resolution. The method distinguishes real object motion from turbulence-induced pseudo-motion and utilizes them differently for restoration. This paper introduces a Thermo-Radiative Motion Coherence Prior, and Motion Reliability-Guided Diffusion to improve temporal information utilization and reconstruction reliability. Extensive experiments demonstrate promising improvements in reconstruction quality and efficiency.

Critical strengths
  1. The paper is interesting and well motivated. Distinguishing beneficial real motion from turbulence-induced pseudo-motion is an intuitive and meaningful direction. The proposed causal-role-aware formulation is reasonable and well aligned with the problem.
  2. The proposed framework is technically sound and well organized. The overall methodology is clearly presented, and the figures are helpful and easy to understand.
  3. The experimental results and ablation studies are convincing. The method achieves strong performance on both benchmarks, with noticeable improvements over existing methods. The availability of the code is also helpful for reproducibility.
Critical weaknesses
  1. The overall framework involves several interacting components, the implementation details of the interacting parts should be explained more clearly (e.g., what is the connection between the disentangled motion features, reliability estimation, and diffusion modulation?).
  2. Although the effectiveness of TRMCP is supported by the experimental results, the respective contributions of its different coherence measurements to motion reliability estimation could be discussed in more detail.
  3. The proposed method achieves favorable inference speed despite its relatively large model size. A clearer discussion of the trade-off between model complexity, restoration quality, and inference efficiency would further strengthen the claim of the paper.
  4. Although the authors provide failure cases under extremely strong turbulence in Appendix A.8, the discussion of the limitations remains relatively brief.
Questions for authors

Please refer to the Weaknesses.

5 comments
AuthorsOpenReview ↗

Response to Reviewer TtVy

We thank the reviewer for the constructive feedback and for recognizing the motivation behind separating object motion from turbulence-induced variations, as well as the strength of our experimental results. We address each point below.

Q1. Connection between motion disentanglement, reliability estimation, and diffusion modulation. We thank the reviewer for this insightful question. Object motion provides cross-frame cues for recovering details, while turbulence-induced pseudo-motion can corrupt those cues. CAMO addresses this by routing the two latent components to different stages of restoration. LMTD estimates both components from latent temporal differences. TCLR uses the turbulence component to rectify the LR latent before diffusion, and MCTE uses the object-motion component to guide the aggregation of neighboring frames after diffusion. Disentanglement determines how each component is used, and reliability determines how strongly the diffusion model updates each region. The estimated object-motion component can still be uncertain under strong turbulence. CAMO estimates this uncertainty, UmotU_{\mathrm{mot}}, from the second-order temporal variation of the object-motion component and the cosine dissimilarity between the two components. In parallel, TRMCP estimates a perturbation score QtherQ_{\mathrm{ther}} directly from the LR frames. MRGD combines the two cues into a reliability map ρ\rho, which controls the strength of DiT feature modulation and scales the predicted velocity (Eqs. 5–7). During training, QtherQ_{\mathrm{ther}} also constrains the relative magnitude of the latent turbulence component through the ratio loss (Eq. 9).

Q2. Roles and turbulence sensitivity of the three TRMCP measurements. We thank the reviewer for this insightful question. TRMCP combines three measurements because local translation compensation can leave different kinds of inconsistencies across frames. The phase residual measures phase changes unexplained by the estimated translation, the radiative residual measures differences in aligned pseudo-radiance, and the edge residual measures changes in aligned gradient magnitudes. We added an analysis across six Fried parameter settings (Table R1). As r0r_0 decreases from 0.140 to 0.025, the mean phase, radiative, and edge residuals all increase monotonically, rising by 92.4%, 91.7%, and 53.2%, respectively. Under fixed normalization scales, the phase and radiative residuals show larger relative changes over the tested range, while the edge residual increases more gradually. The fused score QtherQ_{\mathrm{ther}} rises from 0.04687 to 0.08469. As described in Sec. 3.2, CAMO combines this score with latent motion uncertainty to estimate motion reliability. Table R1 quantifies each measurement’s sensitivity to turbulence strength, while Fig. 7 evaluates the restoration benefit of the fused prior.

Table R1. TRMCP component responses across turbulence strengths. Values are mean ± sample standard deviation across 50 videos at each setting. Normalization scales are held fixed across settings. Smaller r0r_0 indicates stronger turbulence.

r0r_0Phase residualRadiative residualEdge residualQtherQ_{\mathrm{ther}}
0.0250.10881 ± 0.022360.08192 ± 0.019910.06334 ± 0.009870.08469 ± 0.01398
0.0350.09504 ± 0.022080.06936 ± 0.015820.05827 ± 0.008360.07422 ± 0.01169
0.0500.08071 ± 0.021440.05994 ± 0.014630.05426 ± 0.008120.06497 ± 0.01049
0.0710.06914 ± 0.020180.05251 ± 0.015440.05005 ± 0.009070.05724 ± 0.01001
0.1000.06220 ± 0.018490.04738 ± 0.019720.04565 ± 0.011570.05174 ± 0.01078
0.1400.05655 ± 0.016460.04273 ± 0.020050.04134 ± 0.012970.04687 ± 0.01143

Q3. Trade-off between model size, restoration quality, and inference speed. We agree that inference speed should be interpreted alongside model size and restoration quality. CAMO uses the pretrained CogVideoX1.5 backbone for its video prior and controls inference cost through temporal compression and one-step restoration. On 30-frame clips (Table 2), CAMO has 5.804B parameters and runs at 2.322 FPS, about 1.9× faster than the one-step diffusion method DLoRAL (1.224 FPS). CAMO also achieves the highest FasterVQA score among the compared methods (0.3692), while the smaller, non-diffusion FMA-Net achieves higher throughput (5.871 FPS).

Q4. Limitations under extreme turbulence and complex motion. We thank the reviewer for prompting us to clarify CAMO’s limitations. Extreme turbulence can erase fine structures across neighboring frames, making them difficult to recover (Appendix A.8). Occlusions, rotations, and fast nonrigid motion can also leave large residuals after local translation compensation, causing TRMCP to underestimate the reliability of useful motion cues.

Thank the authors for their responses. I have one follow-up question about motion reliability. In Eq. (7), lower reliability reduces the magnitude of the diffusion update. Yet low-reliability regions may also be more severely distorted and require more substantial correction. What is the rationale for applying a smaller update in these regions?

We thank the reviewer for this insightful question. Severely distorted regions may indeed require larger corrections. However, when the temporal cues are unreliable, a larger diffusion update can amplify errors in the predicted correction. In CAMO, TCLR first uses the estimated turbulence component to rectify the LR latent. MRGD then attenuates the contribution of the predicted velocity in low-reliability regions, keeping the preliminary estimate closer to the scaled, rectified latent in Eq. (7). This conservative update limits the influence of uncertain temporal cues, at the cost of potentially retaining residual distortions when TCLR provides insufficient correction.

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

This paper proposes CAMO, a one-step diffusion-based framework for turbulent infrared video super-resolution. It disentangles latent temporal variations into object-motion and turbulence components, which are used to rectify the low-resolution latent and enhance temporal consistency, respectively. Experiments on FLIR-IVSR and M3FD demonstrate promising performance: CAMO achieves the best results across all eight evaluation metrics on FLIR-IVSR and leads on five metrics on M3FD, while remaining competitive on the others.

Critical strengths
  1. The paper provides extensive qualitative results with convincing reconstruction quality.
  2. The paper presents comprehensive quantitative evaluations and comparisons with state-of-the-art methods, demonstrating consistently strong performance across multiple metrics.
Critical weaknesses

Figure 7 evaluates the proposed components mainly through leave-one-component-out comparisons. However, TCLR and MCTE are removed jointly, despite serving distinct purposes: TCLR rectifies turbulence-induced perturbations in the LR latent, whereas MCTE exploits object-motion cues to enhance temporal consistency. This joint ablation does not isolate the individual contribution of either module, nor does it clarify whether their combination provides complementary or synergistic benefits.

Questions for authors

Could the authors report separate ablations of TCLR and MCTE, together with the joint-removal variant, under otherwise identical settings? Additional ablation experiments examining the interactions between LMTD, TRMCP, and the downstream restoration modules would further strengthen the evidence for the proposed design. Furthermore, comparisons with the baseline model would help quantify the performance gains attributable to the proposed components and provide a clearer assessment of their effectiveness.

2 comments
AuthorsOpenReview ↗

Response to Reviewer CnXz

We thank the reviewer for the constructive feedback and for recognizing the quality of our reconstructions and the breadth of our evaluation. We agree that the joint ablation in Fig. 7 leaves the individual contributions of TCLR and MCTE unclear. We have added separate ablations and a same-backbone baseline in Table 3 of Appendix A.10.

Q1. Individual contributions of TCLR and MCTE. We evaluate four configurations: full CAMO, CAMO without TCLR, CAMO without MCTE, and CAMO without both modules. The table also includes the same-backbone baseline discussed in Q3.

ConfigurationPSNR ↑SSIM ↑LPIPS ↓BRISQUE ↓
CogVideoX1.5 baseline31.66420.89560.164251.1095
CAMO w/o TCLR and MCTE35.11800.93180.131044.8297
CAMO w/o TCLR35.96700.93200.121342.2750
CAMO w/o MCTE36.09360.94020.124441.5441
Full CAMO36.73420.94770.116038.0304

Each module improves all four metrics, whether the other module is present or absent. Removing TCLR causes larger drops in PSNR and SSIM than removing MCTE, consistent with its role in rectifying the turbulence-degraded latent before restoration. Removing MCTE causes a larger increase in LPIPS, showing its contribution to perceptual reconstruction quality. The full model performs best across all four metrics, supporting the use of both modules.

Q2. Interactions among LMTD, TRMCP, and the restoration modules. LMTD supplies the turbulence component to TCLR for latent rectification and the object-motion component to MCTE for temporal enhancement. Both components also contribute to the latent motion uncertainty estimate. TRMCP provides an image-domain perturbation cue that is combined with this uncertainty to guide diffusion through MRGD. During training, the TRMCP map also regularizes the relative magnitude of the turbulence component through the loss.

Q3. Comparison with the baseline As shown above, we added a baseline comparison that uses the same CogVideoX1.5 backbone for one-step restoration. Compared with the baseline, full CAMO improves PSNR by 5.07 dB and reduces LPIPS from 0.1642 to 0.1160, with improvements in SSIM and BRISQUE as well. This comparison shows that the CAMO design improves restoration quality.

I appreciate the authors' response. My score remains unchanged.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.