acceptodds
Under review as a conference paper at ICLR 2027

Conditional Residual Decomposition for Distributed Image Compression

Abstract

In distributed image compression (DIC), the compressed target representation and decoder-only side information (SI) jointly determine the reconstruction. However, existing DIC methods do not explicitly characterize the residual that remains after the compressed target representation and SI have both been used. As a result, they do not distinguish the content that can be inferred from SI from the remaining residual that cannot be determined by the available conditions, motivating generative modeling. To achieve this separation, we apply a conditional-expectation projection to the residual left by the compressed target representation, isolating the component predictable from SI from an orthogonal remainder after joint decoding. This orthogonal decomposition yields a variance partition in which the explained fraction defines SI reliability, while the remaining variance determines the minimum MSE of deterministic reconstruction. Building on this decomposition, we propose a DIC method with Conditional Residual Decomposition (CRDIC), which estimates how much the utilized SI reduces the residual left by the compressed target representation and converts this reduction into a spatial estimate of the residual remaining after joint decoding. The deterministic branch uses the SI-predictable residual component to refine the conditional reconstruction, while a stochastic branch models the remaining residual with an amplitude determined by its estimated scale. An antithetic odd projection further suppresses noise-independent corrections in the stochastic branch. Experiments on KITTI Stereo, Cityscapes, and InStereo2K demonstrate superior reconstruction fidelity and perceptual quality across multiple bitrates.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Distributed image compression, side information, conditional residual decomposition, joint decoding

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

This paper studies distributed image compression with decoder-only side information. Using conditional expectation, it decomposes the target-only reconstruction residual into an SI-predictable component and an orthogonal remainder. The proposed CRDIC framework combines SI-assisted deterministic reconstruction with reliability-controlled stochastic residual generation. An odd projection suppresses mean shifts in the stochastic residual, while a control parameter adjusts its amplitude at inference. Experiments on three datasets evaluate the method’s reconstruction fidelity and perceptual quality.

Critical strengths
  1. Good reconstruction quality.

    Experiments on KITTI Stereo, Cityscapes, and InStereo2K demonstrate good reconstruction fidelity and perceptual quality across multiple bitrates.

  2. Detailed supplementary explanations and analyses.

    The appendix provides implementation and training details, together with analyses of residual orthogonality, reliability calibration, and stochastic reconstruction behavior. These additions help clarify the implementation and limitations of the method.

Critical weaknesses
  1. Limited novelty.

    The proposed conditional residual decomposition offers limited novelty. The basic strategy of deterministic reconstruction followed by conditional residual generation already appears in “A Residual Diffusion Model for High Perceptual Quality Codec Augmentation” (Ghouse et al., 2023). The SI reliability estimation and residual-scale control introduced here primarily constitute specific mechanisms within this established framework, limiting the novelty of the overall reconstruction framework.

  2. Poor calibration of the remaining uncertainty.

    The proposed mechanism relies on accurate SI reliability estimates to determine the remaining residual scale through σc=σ01−ρ02\sigma_c = \sigma_0 \sqrt{1 - \rho_0^2}. Overestimating reliability therefore suppresses the estimated residual scale, particularly affecting stochastic restoration when SI is weak. Fig. 6 shows precisely this pattern: reliability is overestimated under weaker SI conditions, while remaining residual scales are generally underestimated.

    The resulting coverage error is substantial. In Fig. 8, even at τ=1\tau = 1, approximately 30–35% of the conditional rank mass lies at each boundary, versus approximately 3% per rank under the uniform reference. This indicates pronounced underdispersion at both evaluated operating points. These results directly weaken the interpretation of the predicted scale and restoration parameter as quantitatively calibrated measures of the remaining uncertainty.

Questions for authors
  1. Presentation of specific contributions.

    Please give greater prominence in the main text and Fig. 1 to the specific contributions in SI reliability estimation and residual-scale control, helping readers distinguish them from existing residual-generation frameworks. The text and image elements in Fig. 1 are also small; enlarging them would improve readability.

  2. Calibration of the remaining uncertainty.

    Figs. 6 and 8 indicate underestimated residual scales and insufficient variation in the generated samples. To what extent do these discrepancies affect the quantitative characterization of the remaining uncertainty and the practical interpretation of the restoration parameter? How do the existing results support the quantitative correspondence between reliability, remaining variance, and stochastic restoration?

2 comments
AuthorsOpenReview ↗

Response to Reviewer ASjX

Thank you very much for taking the time to review. We try our best to address your questions as follows.

W1. Limited novelty relative to DIRAC

Thank you for your comments and suggestions. We have further clarified the novelty and technical contributions of CRDIC as follows.

1. The basic strategy of deterministic reconstruction followed by conditional residual generation already appears in “A Residual Diffusion Model for High Perceptual Quality Codec Augmentation” (Ghouse et al., 2023).

Research problem. DIRAC (Ghouse et al., 2023) addresses perceptual enhancement in single-image compression by generating the residual left by a distortion-optimized base codec. In distributed image comprression (DIC), decoder-only side information (SI) can additionally recover part of the target-only residual. Directly extending residual generation to this setting without distinguishing the contribution of SI may introduce stochastic corrections that overlap with SI-recoverable content, potentially modifying reliably reconstructed details and compromising reconstruction fidelity and semantic consistency. CRDIC addresses this issue by distinguishing the residual predictable from SI from the uncertainty remaining after joint decoding.

Technical distinction. DIRAC combines a distortion-optimized codec with conditional residual diffusion to improve perceptual quality. CRDIC formulates an SI-aware conditional residual decomposition, separating the target-only residual into an SI-predictable component and an orthogonal remainder. The contribution of SI is quantified by reliability ρ2=VS/VT\rho^2=V_S/V_T, which determines the remaining residual scale σc=σ01−ρ02\sigma_c=\sigma_0\sqrt{1-\rho_0^2}. Accordingly, SI-supported information is incorporated into deterministic reconstruction, while stochastic restoration is regulated by the estimated remaining uncertainty. These differences are illustrated in the newly added Fig. 1(f,g) and clarified in the revised Introduction.

2. The SI reliability estimation and residual-scale control introduced here primarily constitute specific mechanisms within this established framework, limiting the novelty of the overall reconstruction framework.

Our contribution lies in the conditional residual decomposition framework for DIC, beyond the individual reliability estimation and residual-scale control modules. By distinguishing the residual recoverable from decoder-only SI from the remaining uncertainty, the proposed formulation provides a unified basis for coordinating deterministic reconstruction and stochastic restoration.

Specifically, the conditional orthogonality yields VT=VS+VT,SV_T=V_S+V_{T,S}, where ρ2=VS/VT\rho^2=V_S/V_T characterizes the fraction of target-only residual variance explained by SI. Accordingly, the remaining variance satisfies VT,S=(1−ρ2)VTV_{T,S}=(1-\rho^2)V_T, establishing a quantitative relationship between SI reliability and the uncertainty remaining after joint decoding. This relationship applies to the conditional reconstruction problem in DIC regardless of the particular codec or generative model employed.

Q1. Presentation of specific contributions

Thank you for your valuable suggestions. Following W1, we have revised the Introduction to clarify our contributions from conventional residual generation, emphasizing SI reliability estimation and residual-scale control. We also added Fig. 1(f,g) to illustrate the technical differences and enlarged the figure elements and labels for readability. All revisions are highlighted in blue.

W2. Poor calibration of the remaining uncertainty.

Thank you for highlighting the calibration discrepancies in Figs. 6 and 8. We acknowledge that the remaining scale is underestimated. The limitations were already discussed in Appendices A.3 and A.4 of the original submission.

In Fig. 6, reliability is evaluated using image-level residual statistics but trained with local error-reduction supervision. Under weak SI, this difference can overestimate ρ0\rho_0 and consequently underestimate σc=σ01−ρ02\sigma_c=\sigma_0\sqrt{1-\rho_0^2}, with additional bias from the learned σ0\sigma_0. In Fig. 8, the odd projection maintains residual centering, while the generated samples exhibit insufficient variation. This partly reflects the use of MSE and LPIPS objectives for reconstruction quality without explicit conditional distribution matching. These explanations are further clarified in the revised Appendix A.5.

These calibration discrepancies reflect the gap between statistical analysis and reconstruction-oriented learning, without affecting the core contributions of CRDIC. The conditional residual decomposition still guides the estimation of SI reliability and the control of stochastic restoration, with the effectiveness of these mechanisms demonstrated by the ablation studies and reconstruction comparisons.

Q2. Calibration of the Remaining Uncertainty

Thank you for your detailed comments and questions. We provide additional experiments to quantify the observed discrepancies and clarify their implications for stochastic reconstruction.

1. To what extent do these discrepancies affect the quantitative characterization of the remaining uncertainty and the practical interpretation of the restoration parameter?

Table 1. Remaining-scale estimation on InStereo2K.

BPPPredicted bias (%)Local-target bias (%)
0.01020-28.52-2.70
0.08062-10.26+2.89
0.15134-0.84+1.42
0.25281+6.27+3.70

(i) Effect on the quantitative characterization of remaining uncertainty.

The remaining scale exhibits larger estimation bias at extremely low bitrates, with relatively small mean bias at higher bitrates. Table 1 compares the predicted scale RMS with the realized residual RMS on InStereo2K. The mean bias is −28.52%-28.52\% at 0.01020 bpp and remains within ±6.3%\pm6.3\% at the two higher bitrates. Replacing the predicted reliability with its local supervision target reduces the absolute mean bias to within 3.7%3.7\% across all four bitrates, indicating that reliability estimation contributes substantially to the low-bitrate scale bias.

We further examine whether improved statistical coverage leads to better reconstruction quality. With the trained model fixed, a global stochastic amplitude multiplier is fitted using conditional-rank statistics. As shown in Table 2, the combined boundary-rank frequency approaches the uniform reference of 6.06%6.06\%, while PSNR and LPIPS deteriorate at all four bitrates. This shows that improving statistical coverage through amplitude rescaling does not necessarily improve reconstruction quality. The difference is consistent with the MSE and LPIPS objectives used to train the stochastic branch.

Table 2. Effect of statistical calibration on reconstruction quality on InStereo2K.

BPPBoundary ranks: Original (%)Boundary ranks: Rescaled (%) →6.06\to 6.06Δ\DeltaPSNR ↑\uparrow (dB)Δ\DeltaLPIPS ↓\downarrow
0.0102059.7513.25-4.553+0.2203
0.0806265.5011.25-7.081+0.2547
0.1513466.5011.00-7.783+0.2870
0.2528162.006.75-7.631+0.2810

(ii) Effect on the practical interpretation of the restoration parameter. The discrepancies affect the absolute stochastic variation produced at a given τ\tau, while its practical role as a relative control of restoration strength remains unchanged. In CRDIC, τ=0\tau=0 corresponds to deterministic reconstruction, and increasing τ\tau progressively increases stochastic restoration around the fixed deterministic result. Its numerical value specifies a restoration level within the learned model, allowing the distortion--perception trade-off to be adjusted without changing the transmitted bitstream or retraining the network.

We have included the additional experiments and clarified the practical interpretation of τ\tau in the revised Appendix A.5. All corresponding revisions are highlighted in blue.

2. How do the existing results support the quantitative correspondence between reliability, remaining variance, and stochastic restoration?

The existing results show that higher SI reliability corresponds to lower remaining residual variance, and that the estimated reliability and residual scale capture this relationship across different SI conditions. Specifically, Fig. 6(a) verifies the approximate orthogonality of the decomposed residual components, while Fig. 6(b) demonstrates a consistent decrease in measured remaining variance as SI reliability increases. Figs. 6(c--e) further show that the predicted reliability and residual scale follow the corresponding measured trends, despite the calibration differences discussed above.

For stochastic restoration, Figs. 5(a,b) demonstrate how varying τ\tau controls the distortion-perception trade-off at fixed bitrates and SI, while Figs. 5(c,d) show the reconstruction responses under different SI conditions. Together, these results demonstrate the empirical correspondence between SI reliability, remaining residual variance, and the resulting stochastic restoration behavior.

Reviewer_SJL4OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This manuscript presents a work on distributed image compression. The main contributions include a formulation of a conditional residual decomposition for distributed image compression, a corresponding conditional residual decomposition strategy. The experiments conducted on different datasets validate the work.

Critical strengths

The paper is well presented. The motivation is clearly described. The experiments are provided with comparisons to the existing methods. The results show the advances in performance gain.

Critical weaknesses

My major concern is on the motivation of this work. What is the target practical application? And how are the experiments on KITTI Stereo, Cityscapes, and InStereo2K representative of the real-world application? It is difficult to be convinced that these experiments instantiate the assumption underlying the proposed distributed setting.

Questions for authors

Please describe the target use case of this work, and explain how the experiments on KITTI Stereo, Cityscapes, and InStereo2K are representative of the real-world application.

5 comments
AuthorsOpenReview ↗

Response to Reviewer SJL4

Thank you for your careful review and for highlighting the importance of the practical applications of our work. We try our best to address your questions as follows.

W and Q. Practical motivation and representative applications.

1. Practical applications. First, in multi-camera surveillance and distributed sensing systems, a central server may already receive and cache a high-quality reference view from a primary camera. Other bandwidth-constrained cameras can independently transmit compressed target images, which are reconstructed using the cached reference as SI. Second, in autonomous driving and multi-view 3D perception, an onboard processing unit may already have a reference image from one camera for visual perception. Other cameras can independently transmit compressed views of the same scene, allowing the processing unit to exploit overlapping scene information without requiring communication between camera encoders.

Rooted in Slepian--Wolf [1] and Wyner--Ziv [2] coding, this distributed setting has been studied in NDIC [3] and ATN [4]. Recent methods, including CODIC [5] and MDIC [6], further demonstrate its potential for bandwidth-efficient image transmission.

We also envision distributed cloud image storage, where a user independently uploads a compressed image and the cloud retrieves related images as SI to assist reconstruction. Correlations among stored images could also be exploited to reduce storage costs, as explored in previous studies [7][8]. Differences in viewpoint and scene content make such correlations challenging to exploit. Existing DIC research provides a foundation for these applications, while the proposed conditional residual decomposition helps characterize the information recoverable from SI and the remaining uncertainty. We have added Appendix A.7, "Practical Applications and Future Directions," to further discuss these applications and future research directions.

2. Representativeness of the experimental setting. Our experiments reflect the multi-camera settings described above. KITTI Stereo and Cityscapes contain stereo views captured in real driving environments, including viewpoint disparities, occlusions, and complex urban scenes. InStereo2K provides stereo pairs captured in indoor environments with diverse object structures and textures. These datasets represent different real-world scenarios in which correlated views can assist reconstruction.

In each dataset, one camera view is independently compressed as the target XX, without access to the reference YY during encoding. The other view is provided exclusively to the decoder as SI. This follows the asymmetric DIC setting evaluated in NDIC [3] and ATN [4], as well as recent generative DIC methods such as CODIC and MDIC. The experiments thus instantiate the separate-encoding and SI-assisted decoding conditions of the target applications.

[1] D. Slepian and J. K. Wolf, “Noiseless coding of correlated information sources,” IEEE Trans. Inf. Theory, vol. 19, no. 4, pp. 471–480, 1973.

[2] A. D. Wyner and J. Ziv, “The rate-distortion function for source coding with side information at the decoder,” IEEE Trans. Inf. Theory, vol. 22, no. 1, pp. 1–10, 1976.

[3] Mital N, Özyılkan E, Garjani A, et al. Neural distributed image compression using common information[C]//2022 Data Compression Conference (DCC). IEEE, 2022: 182-191.

[4] N. Mital, E. Özyılkan, A. Garjani, and D. Gündüz, “Neural distributed image compression with cross-attention feature alignment,” in Proc. IEEE/CVF Winter Conf. Appl. Comput. Vis. (WACV), 2023, pp. 2498–2507.

[5] G. Xu, C. Tan, J. Xiang and J. Zhou, "CODIC: Distributed Image Compression with Distortion-Perception Co-Optimization," in IEEE Transactions on Multimedia, doi: 10.1109/TMM.2026.3718254.

[6] G. Xu, M. Zhang, J. Xiang, C. Tan, Y. Yang, and J. Zhou, “Distributed image compression with multimodal side information at extremely low bitrates,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), pp. 33 662–33 671.

[7] Song X, Peng X, Xu J, et al. Cloud-based distributed image coding[J]. IEEE Transactions on circuits and systems for video technology, 2015, 25(12): 1926-1940.

[8] Bégaint J, Thoreau D, Guillotel P, et al. Region-based prediction for image compression in the cloud[J]. IEEE Transactions on Image Processing, 2017, 27(4): 1835-1846.

The authors tried to provide some examples to address my questions on target applications, but did not introduce any real-world applications or products using DIC. Based on the authors' responses to other reviewers' questions, the authors have better clarified the theoretical contributions. I prefer to recommend "Boardline" (2.5) and still have concerns on the applications.

If the authors could provide a clear evidence of a POC or a real-world product, simply and directly, I would be willing to reconsider my rating. Otherwise, I leave it to the AC to decide whether to treat my score of 2.5 as a 2 or a 3.

AuthorsOpenReview ↗

Official Comment by Authors

Thank you for clarifying your concern. We would like to highlight three implemented and experimentally validated studies of cloud-assisted image coding.

Yue et al. [1] (IEEE TMM 2013) implemented cloud-assisted image coding, where mobile devices transmit compact image descriptors and the cloud reconstructs images using retrieved references. Song et al. [2] (IEEE TCSVT 2015) further developed Cloud-DIC, using syndrome coding and decoder-only SI retrieved from a cloud database, achieving up to 5 dB improvement and 70% bit savings over JPEG with arithmetic coding. Bégaint et al. [3] (IEEE TIP 2018) implemented region-based prediction to exploit correlations among images for cloud storage compression.

These studies provide experimental POC evidence for cloud-assisted distributed image compression, particularly the decoder-only SI setting in Cloud-DIC [2]. Industrial interest is further reflected in Intel's patent [4] (2022) on context-aware image compression, which exploits correlations among photographs to reduce cloud storage costs. The patent was subsequently assigned to Hyundai Motor and Kia in 2024.

[1] H. Yue et al., “Cloud-Based Image Coding for Mobile Devices—Toward Thousands to One Compression,” IEEE TMM, 2013. (https://doi.org/10.1109/TMM.2013.2239629)

[2] X. Song et al., “Cloud-Based Distributed Image Coding,” IEEE TCSVT, 2015. (https://doi.org/10.1109/TCSVT.2015.2416562)

[3] J. Bégaint et al., “Region-Based Prediction for Image Compression in the Cloud,” IEEE TIP, 2018. (https://doi.org/10.1109/TIP.2017.2788192)

[4] Intel, “Context-Aware Image Compression,” US11531850B2, 2022. (https://patents.google.com/patent/US11531850B2/en)

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

This paper studies distributed image compression (DIC): the target image is encoded independently, and the decoder performs joint reconstruction using side information (SI) that is available only at the decoder. The authors point out that existing DIC methods do not explicitly distinguish between "the residual predictable by SI" and "the residual that remains uncertain even after joint decoding." To address this, they propose conditional residual decomposition: given the transmitted representation T, the MMSE residual U = X − E[X|T] under target-only conditions is decomposed via conditional expectation into an SI-predictable component M = E[U|T,S] and an orthogonal remainder E, yielding a variance partition V_T = V_S + V_{T,S}, where SI reliability ρ² = V_S/V_T, and the remaining variance V_{T,S} bounds the deterministic reconstruction error. Based on this, CRDIC is constructed: the SI-predictable component is assigned to the deterministic branch.

Critical strengths
  1. Conditional residual decomposition applies the standard MMSE orthogonal projection theory to DIC, and experimentally verifies the validity of the theory, which is interesting. This is a rare case of "theory not being decorative but directly driving the architecture."

  2. Inference-time controllable perception–distortion trade-off. The scalar τ (Eq.13–14) continuously interpolates between determinism (τ=0) and full residual (τ=1), and Fig.5 verifies the expected behavior: when τ increases to 0.2, LPIPS decreases by 0.0146–0.0382, while PSNR only loses 0.065–0.077 dB.

  3. Strong results, efficient, and honest disclosure of limitations. CRDIC leads in both fidelity and perceptual metrics across three datasets against all ten baselines, with the encoder having only 2.75M parameters (0.0033 s), and decoding is 18.7–23.2× faster than diffusion-based DIC.

Critical weaknesses
  1. The gap between theory and model is real, weakening the accuracy of the decomposition. The ideal sample E′∼p(E∣T,S)E' \sim p(E|T,S) is approximated by a learned, amplitude-constrained stochastic residual, and the paper acknowledges that it does not fully reproduce the conditional distribution (A.4). More specifically, the predicted residual scale σc\sigma_c is systematically lower than the measured residual RMS (Fig. 6e, PCC 0.67–0.83 but biased low), and the reliability prediction ρ\rho has moderate accuracy (Fig. 6c, PCC 0.57–0.65, and is systematically overestimated). Since ρ\rho drives σc\sigma_c and the deterministic/stochastic allocation, this calibration bias means that the theoretical precise variance control (Eq. 11–12) only holds approximately in the actual model.

  2. The side information here appears to use lossless stereo view images: this is too idealized for general applications, because in most cases the receiver receives stereo image pairs that are both lossy. The paper should investigate the enhancement effect of side information on the main view under different degradation scenarios.

  3. The comparison baselines and references lack some newer perception-based multi-view compression pipelines [1]. The authors should compare against some single-view image compression schemes, such as VVC and [2][3], as in previous works. Since the authors also compare on perceptual quality, some perceptual visual compression schemes should be added.

[1] Xia Y, Zhou Y, Li Z, et al. SDiD: Shared Diffusion Prior for Efficient Distribute Stereo Image Compression[C]//Forty-third International Conference on Machine Learning. 2026.

[2] Wei Jiang, Jiayu Yang, Yongqi Zhai, Feng Gao, and Ronggang Wang. Mlic++: Linear complexity multi-reference entropy modeling for learned image compression. ACM Transactions on Multimedia Computing, Communications and Applications, 21(5):1–25, 2025.

[3] Dailan He, Ziming Yang, Weikun Peng, Rui Ma, Hongwei Qin, and Yan Wang. Elic: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 5708–5717. IEEE, 2022.

Questions for authors
  1. Verify the enhancement effect of side information on main-view decoding across different degradation scenarios.

  2. Add some single-view image compression baselines, including traditional schemes (VVC), LIC-based [1][2], and generative-based [3][4].

[1]Wei Jiang, Jiayu Yang, Yongqi Zhai, Feng Gao, and Ronggang Wang. Mlic++: Linear complexity multi-reference entropy modeling for learned image compression. ACM Transactions on Multimedia Computing, Communications and Applications, 21(5):1–25, 2025.

[2]Dailan He, Ziming Yang, Weikun Peng, Rui Ma, Hongwei Qin, and Yan Wang. Elic: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 5708– 5717. IEEE, 2022.

[3]Xia Y, Zhou Y, Wang J, et al. Towards efficient low-rate image compression with frequency-aware diffusion prior refinement[C]//Proceedings of the AAAI Conference on Artificial Intelligence. 2026, 40(13): 10942-10950.

[4]Zhang T, Luo X, Li L, et al. Stablecodec: Taming one-step diffusion for extreme image compression[C]//2025 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2025: 17379-17389.

4 comments
AuthorsOpenReview ↗

Response to Reviewer mejc

Thank you very much for taking the time to review and for your support. We try our best to address your questions as follows.

W1. Gap between theoretical decomposition and learned reconstruction.

Thank you for pointing out this limitation. As you noted, we have discussed this limitation in Appendix A.4. We would like to provide further clarification.

In Fig. 6, reliability is supervised by local reconstruction-error reduction but evaluated against image-level residual statistics. This difference contributes to reliability overestimation and the resulting scale bias, together with errors in the learned σ0\sigma_0. Moreover, the stochastic branch is optimized using MSE and LPIPS for reconstruction quality, which does not explicitly enforce the full conditional residual distribution and contributes to the underdispersion observed in Fig. 8.

These calibration discrepancies affect the numerical accuracy of uncertainty estimation, but do not negate the core contribution of conditional residual decomposition in guiding deterministic reconstruction and stochastic restoration. Fig. 6(a,b) demonstrates the expected residual relationships, while the ablation studies and Fig. 5 validate the proposed mechanisms and controllable distortion-perception reconstruction. We have further quantified these discrepancies and clarified their implications in the revised Appendix A.5.

W2 and Q1. Practical assumptions and reconstruction robustness under degraded SI.

Thank you for this valuable comment. Our asymmetric DIC setting considers scenarios where a high-quality reference image is already available at the decoder, as in previous studies on decoder-side SI [1][2]. For example, in centralized multi-camera systems, a reference view may be received through a dedicated high-bandwidth connection and cached at the processing unit. Other cameras independently transmit compressed target views, which are reconstructed using the cached reference. We have further clarified these practical applications in Appendix A.7.

We appreciate the reviewer's suggestion to further examine degraded SI, as this provides additional insight into the practical robustness of our method. Table 1 and Fig. 5(c,d) of our original manuscript already demonstrate the benefits of SI-assisted reconstruction under varying SI conditions. Following this suggestion, we have added a more detailed evaluation in Appendix A.4, with the new content highlighted in blue. The new Fig. 9 compares CRDIC with MDIC [3], CODIC [4], and NDSC [2] under absent, mismatched, shifted, blurred, and noisy SI on InStereo2K. We fix the model and target bitrate conditions and set τ=0\tau=0 to evaluate SI-assisted deterministic reconstruction.

Compared with absent SI, matched SI improves CRDIC by 4.13 dB in BD-PSNR and 0.282 in BD-LPIPS gain. Under Gaussian blur (σ=2\sigma=2), CRDIC still achieves improvements of 0.55 dB and 0.070, respectively. It also outperforms NDSC, CODIC, and MDIC by 3.17, 6.39, and 6.90 dB in BD-PSNR under this condition. These results demonstrate that CRDIC maintains reconstruction benefits under moderate SI degradation and quantify its performance across different degradation levels.

[1] Mital N, Özyılkan E, Garjani A, et al. Neural distributed image compression using common information[C]//2022 Data Compression Conference (DCC).

[2] Whang J, Nagle A, Acharya A, et al. Neural distributed source coding[J]. IEEE Journal on Selected Areas in Information Theory, 2024.

[3] G. Xu, M. Zhang, J. Xiang, C. Tan, Y. Yang, and J. Zhou, Distributed image compression with multimodal side information at extremely low bitrates, CVPR, 2026.

[4] G. Xu, C. Tan, J. Xiang and J. Zhou, CODIC: Distributed Image Compression with Distortion-Perception Co-Optimization, in IEEE Transactions on Multimedia, doi: 10.1109/TMM.2026.3718254.

AuthorsOpenReview ↗

Response to Reviewer mejc

W3 and Q2. Add some single-view image compression baselines.

Thank you for your valuable suggestions. Following your recommendation, we have added comparisons with representative single-view compression methods, including VVC, MLIC++, ELIC, and StableCodec. We also attempted to include DiffCR, but its official repository only provides inference code without the complete training implementation, making a fair reproduction under our experimental protocol difficult. Therefore, we selected AEIC [5], a recent perceptual image compression method, as an alternative. The additional results are presented in the experiments section (Fig. 3 and Fig. 10), with the new content highlighted in blue.

We also appreciate the reviewer's recommendation to compare with SDiD, an excellent recent work on stereo image compression. Its official code was released in October 2026, but the complete training implementation is not yet available. Differences in dataset preprocessing and evaluation metrics also prevent us from directly using its reported results. We apologize for being unable to provide a fair comparison within the rebuttal period and have added SDiD to the Related Work section. Meanwhile, our original manuscript already compares with two contemporary perception-oriented DIC methods, MDIC (CVPR 2026) and CODIC (IEEE TMM 2026). As shown in Fig. 3, CRDIC consistently outperforms both in distortion and perceptual quality across all three datasets.

Table 1: BD-Metric comparison of single-image compression methods relative to CRDIC. Negative BD-PSNR and positive BD-LPIPS/BD-DISTS indicate worse performance than CRDIC. Arrows indicate the preferred direction of each metric.

MethodBD-PSNR (dB) ↑BD-LPIPS ↓BD-DISTS ↓
KITTI Stereo
CRDIC (Ours)0.000.0000.000
MLIC++-5.77+0.321+0.169
VVC-4.45+0.357+0.213
ELIC-3.64+0.245+0.129
StableCodec-9.96+0.306+0.151
AEIC-6.94+0.196+0.073
Cityscapes
CRDIC (Ours)0.000.0000.000
MLIC++-6.67+0.331+0.199
VVC†-7.81+0.520+0.320
ELIC-5.55+0.295+0.183
StableCodec-10.82+0.305+0.143
AEIC-10.40+0.195+0.086
InStereo2K
CRDIC (Ours)0.000.0000.000
MLIC++-4.36+0.304+0.173
VVC-5.05+0.330+0.209
ELIC-3.35+0.254+0.147
StableCodec-10.50+0.213+0.082
AEIC-9.34+0.203+0.098

We compute the average performance differences between CRDIC and each single-image compression baseline from their rate-distortion and rate-perception curves over the common bitrate intervals (Fig. 3 and Fig. 10 in the revised manuscript). As shown in Table 1, CRDIC consistently achieves higher PSNR and lower LPIPS and DISTS across all three datasets. These improvements reflect the benefit of exploiting decoder-only SI, which is unavailable to single-view codecs. Among the baselines, generative methods show stronger perceptual performance, with AEIC exhibiting smaller LPIPS and DISTS gaps than ELIC across all three datasets. These results further demonstrate the effectiveness of SI-assisted reconstruction in both distortion and perceptual quality.

[5] Zhang T, Liu D, Chen C W. Ultra-low bitrate perceptual image compression with shallow encoder, CVPR. 2026.

Thank you for the detailed response from the authors. The authors have added substantial new experiments, which effectively validate the reliability of the proposed method. I keep my original score and lean towards accepting this paper.

AuthorsOpenReview ↗

Official Comment by Authors

Thank you so much for your kind words and support! We are glad that the additional experiments have helped validate the reliability of our method and address your concerns. We will continue working to further improve our research and future work.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to All Reviewers

We sincerely thank all reviewers for their careful evaluations, constructive comments, and valuable suggestions. We have carefully revised the manuscript and conducted additional experiments to address the concerns raised. The main improvements and clarifications are summarized below.

Key Improvements and Clarifications

  1. Technical Novelty and Presentation. Clarified the distinction between CRDIC and conventional residual generation methods, particularly DIRAC. Revised the Introduction to emphasize the conditional residual decomposition. Added Fig. 1(f,g) and improved its readability. Asked by Reviewer ASjX.

  2. Extended Experimental Comparisons. Added single-view compression baselines, including VVC, MLIC++, ELIC, AEIC, and StableCodec (Figs. 3 and 10). Expanded the Related Work section to cover recent methods, including SDiD. Asked by Reviewer mejc.

  3. Robustness. Added experiments on InStereo2K under absent, mismatched, shifted, blurred, and noisy SI, with comparisons against MDIC, CODIC, and NDSC (Appendix A.4). Asked by Reviewer mejc.

  4. Practical Applications. Further clarified the practical motivation and representative applications of asymmetric DIC (Appendix A.7). Asked by Reviewer SJL4.

  5. Theoretical Approximation and Calibration. Clarified the differences between theoretical conditional statistics and learned network predictions. Added experiments examining residual-scale estimation and statistical recalibration, together with their effects on reconstruction quality (Appendix A.5). Asked by Reviewer ASjX.

All revisions are highlighted in blue in the updated manuscript. Please see our detailed responses below each review for further explanations and experimental results. We sincerely appreciate the reviewers' feedback and welcome any additional questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.