acceptodds
Under review as a conference paper at ICLR 2027

LoopMoEVR: Loop-Based Degradation-Aware Mixture-of-Experts for Unified UHD Video Restoration

Abstract

Recently, unified high-definition image restoration has attracted considerable attention; however, existing models tend to excessively increase their depth in pursuit of improved generalization, which often yields only limited gains. Meanwhile, loop-based learning paradigms have drawn widespread attention due to their low parameter counts and strong regression capability, as exemplified by GPT-6 and looped Transformers. In this paper, we introduce the loop learning paradigm to address restoration tasks that require cross-domain learning. Specifically, we propose LoopMoEVR, a loop-based mixture-of-experts model capable of handling degraded ultra-high-definition (UHD) inputs. First, a degradation-conditioned low-rank loop embedding is designed to construct input-dependent stage conditions. Second, a spatio-temporal iterative adaptive normalization module, termed IterAda3DN, is developed to fuse local features with global loop context, thereby performing position-wise affine modulation. Finally, the expert branches further integrate the attention-updated local and global video states with the loop conditions to generate dedicated modulation parameters, while an input-conditioned depth predictor adaptively configures the number of loop iterations. With only approximately 0.884M trainable parameters, the proposed model uniformly handles UHD video dehazing, deraining, denoising, and low-light enhancement tasks, achieving state-of-the-art restoration performance on both public benchmarks and real-world scenarios.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Video Restoration, Moe, UHD video, Loop

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_mm9iOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

Setting the eyes on the efficiency, this paper addresses the tasks on unifying 4K video dehazing, deraining, denoising, and LLE. Named as LoopMoEVR, the proposed method reuses a video-processing backbone across iterations, combines degradation information with iteration embeddings through a low-rank interaction, and uses local features, pooled video states, and cross-iteration momentum to condition normalization and updates. The paper also designs a separately trained predictor to select the number of iterations for better dynamic efficiency. Experiments cover different datasets and degradation types. Component ablations and qualitative experiments on real-world images are also included within the paper.

Critical strengths
  1. Clear motivation and a useful idea on the tradeoff of quality and parameter. The conditioning settings are tested and confirmed.
  2. Extensive ablations on adaptive depth and temporal visuals. The results also support input-dependent iteration allocation while keeping stability.
  3. Qualitative comparison is included within the paper.
Critical weaknesses
  1. While promoting the efficiency of this domain is applaudable, the problem is, from Tab. 1, also from the design of the proposed method, the reduced params are from the trade-off of computation speed, i.e., by iterations. Compared to others, the problem is more explicit, e.g., 0.51 s v.s. 0.48 s. This renders the Fig. 1 misleading. As less params will not contribute to faster processing speed and not suitable for certain mobile phone uses/implementations. Justifications are needed regarding this claims.
  2. The generalization claim is too broad in reviewer's minds. Real-world evidence only consists of two sets of still images, while non-UHD PSNR falls below on both tested datasets. These results do not support the generalization.
  3. While Tab. 6 compares dense and expert branches, the current presentation is without reporting their params or computation needs. Adding shared experts improves PSNR by 1.03 dB, but also adds capacity. Are the gains from a comparable parameter or computation budget?
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.

A point regarding methodology part shall also be noted, as that, the current equations and notations are rambling, forcing the readers to read the paper with extreme attention. The notations are also less connected to the presented figures.

Reviewer_BAEeOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes LoopMoEVR for unified 4K video dehazing, deraining, denoising, and low-light enhancement. The method repeatedly refines video features through a mixture-of-experts network with parameters shared across iterations, enabling multi-stage restoration with fewer parameters. Its main designs include low-rank conditional encoding jointly driven by degradation and iteration, adaptive normalization that integrates local spatiotemporal features with cross-iteration states, and state-aware expert modulation. An input-dependent depth predictor selects three to five iterations to balance restoration quality and computational cost. The authors report the highest average PSNR on UHV-4K-AIO with approximately 0.884M parameters. They further assess the method’s effectiveness and scope through ablations on parameter sharing, condition fusion, and dynamic depth, alongside evaluations of time-varying degradations, temporal consistency, non-4K benchmarks, and qualitative results on real-world images.

Critical strengths
  1. With the same network width and four iterations, parameter sharing reduces the model size from 3.116M to 0.884M parameters, a reduction of approximately 71.6%, while PSNR decreases by only 0.08 dB. The full model also achieves the highest average PSNR among the methods listed on UHV-4K-AIO. These results provide direct evidence of the parameter efficiency of the looped architecture with shared weights, demonstrating its value for compact, unified UHD video restoration models.

  2. With the depth fixed at four iterations, the full method outperforms a LoopMoE-style video baseline by 1.20 dB. The dynamic depth strategy also outperforms a random strategy with the same distribution of iteration counts within each task by 0.51 dB. The condition fusion experiments additionally control the parameter budget of the condition generator. These comparisons help identify the incremental contributions of video conditioning and input-dependent computation allocation, strengthening confidence in the method’s effectiveness.

Critical weaknesses
  1. The substantive contribution of the new designs over LoopMoE requires stronger controlled comparisons. Looped parameter sharing, sparse experts, and modulation conditioned on iteration and current state all have direct precedents. The main additions are degradation conditioning, global video information, cross-iteration states, and depth prediction. Although the ablations support their effectiveness, the specific architectures, parameter counts, and computational costs of the LoopMoE-style video baseline and dense alternative branches are insufficiently described. It is therefore difficult to determine how much of the improvement comes from the proposed mechanisms rather than differences in conditioning branch capacity or resource allocation.

  2. Missing implementation and training details affect the reproducibility of the core results. The appendix only outlines auxiliary losses, such as mask constraints and expert overlap constraints, without providing complete definitions or specific weights. The normalization schemes and weights for runtime and memory costs used to generate depth supervision labels are also unspecified. Other details, including the network’s channel configuration and the exact rank of the low-rank conditional encoding, remain incomplete. These settings directly affect restoration performance, expert behavior, and predicted iteration counts, particularly the trade-off between quality and cost achieved by dynamic depth. The current description is therefore insufficient to reliably reproduce the model or assess how sensitive the reported gains are to these settings.

Questions for authors
  1. Are the resource configurations of the baselines and the full model comparable? Please provide the specific architectures, parameter counts, and computational costs of the LoopMoE-style video baseline and the Compact Dense and Expanded Dense alternatives. Which conditioning branches are retained or modified relative to the full model? This information would help determine whether the gains arise from the proposed mechanisms or differences in capacity and resource allocation.

  2. Can the key definitions and configurations required for reproduction be provided in full? Please specify how auxiliary losses, including the mask and expert overlap constraints, are computed, and provide all loss weights, the network’s channel configuration, and the exact rank used in the low-rank conditional encoding. Are these settings consistent across the ablation experiments?

  3. How exactly are the supervision labels for dynamic depth generated? In the quality–cost objective J(K)J(K), how are runtime and GPU memory costs measured and normalized, and how are their weights determined? Are these costs measured at the training crop size or on full 4K inputs? These choices directly affect the target iteration counts and the resulting trade-off between quality and speed.

  4. How should the latency change after parameter sharing be interpreted? In Table 8, parameter sharing substantially reduces model size but increases runtime from 515 ms to 560 ms. Do both configurations use the same hardware, numerical precision, and inference implementation, and are the timings based on repeated measurements? What do the authors consider the main source of the additional runtime, and do the parameter savings also reduce peak GPU memory usage?

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

This paper proposes LoopMoEVR, a parameter-shared looped mixture-of-experts model for unified UHD video restoration (dehazing, deraining, denoising and low-light enhancement). It adapts LoopMoE from language modeling to video with three additions: a degradation-conditioned low-rank loop embedding, a spatio-temporal iterative adaptive normalization (IterAda3DN) that carries a clip-level momentum state across iterations, and state-aware expert modulation. A depth predictor, trained in a separate stage, chooses 3–5 iterations per input. With 0.884M parameters, the model reports the highest mean PSNR on UHV-4K-AIO. Additional experiments cover time-varying degradations, temporal consistency, non-UHD datasets, a user study and ablations.

Critical strengths

Very compact and fast for native 4K video: 0.884M parameters and ~0.5s per window, with mean PSNR above MoCRA at about a quarter of its parameters.

The ablations are fairly thorough: components, conditioning form, expert branch, condition order, and parameter sharing.

Critical weaknesses

Several reported numbers are implausible or internally inconsistent. In Tab. 3, multiple baselines score much higher under the harder time-varying setting than in the static one. In Tab. 8, the shared and unshared models report different latencies at identical compute.

The SOTA claim is narrow. It rests on mean PSNR on a single small benchmark. AirNet is better on mean SSIM and LPIPS, BasicVSR++ clearly wins on non-UHD data, and the paper does not say whether the baselines were retrained.

Novelty is limited. The method is largely a transfer of LoopMoE plus established components, and the dynamic depth predictor saves only 8.9% of inference time.

Questions for authors

Why do several baselines score higher under time-varying degradations than in the static setting? How was that protocol implemented?

Were all baselines retrained on UHV-4K-AIO with the same budget? Were the runtimes measured on the same hardware and resolution?

Why does the shared configuration run slower than the unshared one at identical compute?

Please report FLOPs and peak memory at 4K, plus results with multiple seeds and full-clip (not center-frame) evaluation.

How do you reconcile the claim of effectiveness at non-UHD resolutions with BasicVSR++ clearly outperforming LoopMoEVR?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.