acceptodds
Under review as a conference paper at ICLR 2027

INTERPRETABLE IMAGE RESTORATION VIA MINIMUM DESCRIPTION LENGTH ROUTING

Abstract

All-in-one image restoration must balance model capacity, computational cost, and the disentanglement of degradations. Conventional Mixture-of-Experts routing relies on heuristic probabilities, producing opaque decisions and potential expert collapse.Grounded in the Minimum Description Length (MDL) principle, we propose MDLE(Minimum Description Length-routed Experts), an interpretable restoration framework that selects experts by jointly minimizing the coding length comprising model complexity and data-fitting error.Its Asymmetric Heterogeneous Expert Pool combines FFT-Attention for global frequency correction with spatial experts for local structure preservation, enabling degradation-aware on-demand computation. Experiments on restoration benchmarks show competitive reconstruction quality and efficient inference. Ablations further demonstrate that routing decisions correlate with degradation types, providing an information-theoretic account of expert selection rather than an empirical black-box gate.Although the quantitative gains over existing multi-routing approaches are currently modest, MDLE offers a distinct perspective on image denoising, dehazing, and other restoration tasks by grounding routing in an explicit information-theoretic objective. This formulation may provide a foundation for further advances and help establish a new direction for interpretable, degradation-aware image processing.

Primary area:
interpretability and explainable AI
Keywords:
All-in-One image restoration, Mixture-of-Experts routing, Minimum Description Length, Interpretable expert routing

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_U2mgOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

All in one restoration is important within vision community. To advance, this paper studies the design of expert selection modules. Named as MDLE, the model uses a multiscale encoder-decoder arch and a pool of attention, CNNs, and MLPs experts, with a router combining different terms and scores to select and fuse a subset of experts. Training uses a set of losses and performance on three tasks is reported. Additional results on deblurring and low-light enhancement are also included. Ablations on factors and visual examples are provided.

Critical strengths
  1. Intuitive and intended routing. Clear motivations on this. A concrete architectural setting is provided for the expert pool.
  2. Extensive and informative ablations on the proposed method, adding credibility to this paper. Also the performance is promising.
  3. The paper is easy to follow. Figures, tables and formatting are good and accessible for readers.
Critical weaknesses
  1. Justifications are needed regarding the training datasets and baseline sources. Currently these details are not specified. Also, a point is that, the proposed model scores 34.15 dB in Tab. 1 but 34.42 dB in Tab. 3. Any different settings on this?
  2. The reviewer finds that, since the eye of this paper in on the interpretability, but the measurements are not that enough or thorough. While Fig. 5 shows the activation frequencies, such a difference may still from the learning, the association alone does not establish the claimed <physical decoupling> of this paper.
  3. Please specify or add ablations on the objective of MDL, since it uses log parameter count, but the router also adds learned scores. Consider separating the gains from these terms.
Questions for authors

See Weaknesses. The reviewer's rating of 2 is mainly based on these. When with satisfactory rebuttal, the reviewer is happy to adjust the score.

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

MDLE is an all-in-one restoration network built on a multi-scale encoder and decoder with a mixture of three heterogeneous experts (FFT-Attention, ConvNeXt, SparseMLP). Routing is described as minimizing an MDL two-part code. The “data term” is log₂ of the MSE of a low-rank projection Wₖᵀ Wₖ x of the input, and the “model term” is log₂ of the expert’s parameter count. The score adds learned linear terms on pooled spatial and frequency features, and passes through a softmax with a load-balancing loss. The paper reports large gains on the three-task benchmark (average 34.15 vs 32.73 dB), GOPRO deblurring (+3.2 dB over MoCE-IR) and LOL. It also includes a 2×2 ablation of the router and the expert pool.

Critical strengths
  1. The motivation, more principled and inspectable MoE routing for restoration, is reasonable.
  2. The 2×2 factorial ablation separating the router from the composition of the expert pool is the right experimental design in principle.
Critical weaknesses
  1. The routing rule is not MDL in any meaningful sense. The data term measures how well the degraded input lies in each expert's low-rank linear subspace (x̂ₖ = Wₖᵀ Wₖ x). It does not measure the code length of the restored output or of the clean target, which is what selecting among restoration models would require. The model term log₂|θₖ| is a constant for each expert, and the difference between experts is only about one bit, so it acts as a fixed bias. Together with the learned linear gates gₖ and hₖ, the router reduces to a standard learned gate with an added subspace-residual feature. The claims of "rigorous", "mathematically complete" and "physically grounded" routing are therefore unsupported. Computing the data term also requires evaluating every expert's projection, which undercuts the claimed sparse, on-demand computation, and Table 3 indeed shows essentially unchanged FLOPs across all variants.

  2. The reported results are implausible and internally inconsistent.

    • On BSD68 at σ = 25, MDLE reports 33.41 dB, about 2 dB above every all-in-one baseline and above strong dedicated denoisers. At σ = 50 it is worse than every baseline in PSNR (27.98 dB) while its SSIM (.865) is far higher.
    • On SOTS, PSNR improves by 2.6 dB over DSwinIR while SSIM decreases.
    • 33.32 dB on GOPRO from an 11M all-in-one model would exceed most dedicated deblurring networks, and the paper does not state whether GOPRO and LOL models were trained separately or jointly.
    • The abstract describes the gains as "currently modest", which contradicts the 1.4 to 3.2 dB improvements reported in the tables.
  3. Several references appear incorrect or fabricated. AirNet is attributed to Chen, Chu, Zhang and Sun (the NAFNet authors), while Li et al. (2022) is cited separately under the same title. The all-in-one IDR baseline is cited as a 2024 self-supervised denoising paper with a different author list. MoCE-IR is cited as "Zhang, Li et al., arXiv 2024", while the actual MoCE-IR paper (Zamfir et al., CVPR 2025) appears separately as "Complexity Experts". DSwinIR carries the placeholder identifier arXiv:2401.00000. AdaptIR is described as a heterogeneous frequency-spatial MoE with an author list that does not match the published work. "SwinV2-MoE" points to an expert-parallelism systems paper, DSMoE is a language model paper presented as a restoration method, and M2IR in Table 2 has no reference at all. Together these cast serious doubt on the integrity of the comparison tables.

  4. The claimed applications exceed the evidence. The paper repeatedly invokes wavefront correction, phase bias, astigmatism, metamaterial imaging and sub-wavelength processing errors, yet no experiment involves such degradations. The downstream evaluation (detection, segmentation, text detection, pose estimation) and the real-camera results are qualitative only, with no task metrics comparing degraded inputs, baseline restorations and MDLE restorations.

Questions for authors

N/A

Reviewer_PNrXOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes MDLE, a mixture-of-experts framework for all-in-one image restoration. It combines FFT-Attention, ConvNeXt, and SparseMLP experts with a routing score motivated by Minimum Description Length. The score balances low-rank reconstruction error and expert parameter count, supplemented by learned spatial and frequency signals. The paper evaluates the method on several restoration tasks and uses ablations and routing visualizations to examine the contributions of the router and expert pool.

Critical strengths
  1. The paper addresses a useful problem: allocating model capacity according to the restoration needs of different inputs. Including explicit reconstruction and complexity terms gives the routing mechanism components that can be inspected and analyzed.
  2. The 2×22\times2 ablation in Table 3 is helpful. Comparing both routing strategies with both expert pools provides evidence for the contributions of the complete routing scheme and expert heterogeneity. This is more informative than comparing only the final model with external baselines.
  3. The evaluation covers denoising, deraining, dehazing, deblurring, and low-light enhancement, with both quantitative results and visual examples. Table 1 reports a higher average PSNR than the listed baselines. The inclusion of runtime and FLOPs in Table 3 also helps assess the quality–computation trade-off.
Critical weaknesses
  1. The choice of experts needs more justification. Why were FFT-Attention, ConvNeXt, and SparseMLP chosen? It is unclear whether this combination follows a specific design principle or was selected through manual experimentation. The comparison with an all-FFT expert pool is useful, but does not establish whether other expert choices or combinations could achieve comparable or better results.
  2. Expert specialization and the added value of routing need more direct evidence. The routing heatmaps show which tasks are assigned to each expert, but do not directly demonstrate that each expert performs better on its intended degradations. It is also unclear whether a deeper shared network, or a network combining frequency and spatial modules without routing, could integrate these capabilities under a comparable computational budget. Such controls would help establish what expert routing adds.
  3. The contribution of the MDL score itself remains unclear. Table 3 compares different routers and expert pools and provides evidence that the full routing scheme improves performance. However, this scheme includes reconstruction error, a parameter-count penalty, learned spatial and frequency scores, and an auxiliary loss. With the expert design held fixed, further ablations are needed to determine how much of the improvement comes from the MDL score itself and how much comes from the other scoring or training components.
Questions for authors

As described in "Critical Weaknesses". I would considering the final rating after authors' rebuttal.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.