acceptodds
Withdrawn from ICLR 2027

Towards All-in-One Image Restoration: A Sparse Mixture-of-Experts Framework with Task-Adaptive Routing

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Image restoration aims to reconstruct high-quality visual content from degraded observations caused by blur, noise, haze, or resolution loss. While recent deep learning approaches have achieved remarkable performance on individual restoration tasks, most existing methods still follow a task-specific paradigm that trains separate models for different degradations. This design limits their practicality in real-world scenarios where degradations are often unknown, mixed or spatially varying. In this paper, we propose MoERestore, a unified image restoration framework that integrates task-adaptive sparse mixture-of-experts routing with degradation-aware conditioning. The proposed architecture dynamically routes each input to a small subset of lightweight experts according to degradation characteristics, enabling task-adaptive capacity allocation while alleviating cross-task interference. To provide explicit guidance for restoration under ambiguous or mixed degradations, we introduce a degradation-aware module that leverages a vision-language model to generate semantic descriptions of corruption patterns. These descriptions are encoded as conditioning signals to guide the restoration process. In addition, a high-frequency feature extraction branch is incorporated to enhance fine-grained texture reconstruction. Extensive experiments on super resolution, deblurring, dehazing, and denoising benchmarks demonstrate that MoERestore achieves strong and consistent performance across diverse restoration scenarios and generalizes well to challenging real-world degradations

Primary area:
generative models
Keywords:
Computer vision, Image Restoration

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_QkL8OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes MoERestore, an all-in-one image restoration framework that integrates task-adaptive sparse mixture-of-experts routing with degradation-aware semantic guidance. Specifically, the method introduces sparse expert routing to alleviate cross-task interference, a vision-language model to provide degradation-aware conditioning, and a high-frequency feature extraction branch to enhance texture reconstruction.

Critical strengths
  1. Integrating sparse mixture-of-experts routing with degradation-aware semantic guidance provides a reasonable approach to handling diverse restoration tasks.

  2. The proposed framework combines task-adaptive expert routing, vision-language guidance, and high-frequency feature extraction to improve restoration performance.

Critical weaknesses
  1. The methodological novelty appears relatively limited. Mixture-of-Experts (MoE) architectures have already been extensively explored in all-in-one image restoration, such as FoundIR, MoCE-IR, and MiM-DiT: MoE in MoE with Diffusion Transformers for All-in-One Image Restoration. The paper does not sufficiently clarify its fundamental technical differences and advantages over these representative MoE-based restoration methods.

  2. The experimental comparisons under composite degradations are insufficient. More comparisons with recent state-of-the-art methods specifically designed for composite degradation restoration are limited.

  3. The ablation studies are not sufficiently comprehensive. Table 4 only reports ablation results on the GoPro deblurring dataset, which cannot adequately demonstrate the contributions of the proposed components to all-in-one restoration. In particular, the effectiveness of sparse expert routing and degradation-aware semantic guidance under different restoration tasks and mixed degradation scenarios remains unclear.

Questions for authors

See the above Weaknesses parts.

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

This paper proposes MoERestore, an all-in-one image restoration framework built on a latent-space MMDiT backbone. It incorporates a sparse mixture-of-experts architecture into the MLP layers, combining shared, routed, and random experts. The framework also uses degradation descriptions generated by Qwen2.5-VL for semantic conditioning and a high-frequency feature branch to support detail reconstruction. The authors evaluate the method on super-resolution, deblurring, dehazing, and denoising tasks, supplemented by real-world evaluations and component ablation studies. They report competitive performance across these restoration scenarios.

Critical strengths

The paper addresses unified image restoration under unknown and mixed degradations, which has practical value. It combines dynamic expert allocation with semantic conditioning to meet the restoration requirements of different tasks, providing complementary support for task adaptation, degradation awareness, and detail reconstruction. The overall approach is clear and well motivated. The method achieves competitive results across multiple image restoration tasks, and the module ablations further demonstrate that the MoE, semantic conditioning, and high-frequency branch improve restoration performance.

Critical weaknesses
  1. The paper contains some grammatical or spelling errors. In addition, some interpretations of the results are inconsistent with the data in the tables. For example, “a schematic diagram of the Spare-MoE” in Figure 1 should be corrected to “a schematic diagram of the Sparse-MoE”; Section 4.6 contains “Overall, … Overall, …”; and in the experimental analysis, the PSNR drop from removing the high-frequency branch is 3.89 dB, which is greater than the 3.59 dB drop from removing MoE, contradicting the analysis in the text.

  2. Some methodological and experimental details are insufficient. For example, how is the approach integrated into MMDiT, is standard joint text–image attention retained, and does the semantic information from Qwen-VL affect expert routing, and if so, how? If not, have the authors considered using semantic information and features together to select experts, making the experts more degradation-specific and semantically consistent? Furthermore, the methods compared are mainly from 2022–2025, and comparisons with or discussions of the latest methods from 2026 could be added.

  3. The paper could include an end-to-end complexity and efficiency analysis, such as parameter count, FLOPs, memory usage, or inference time. It is recommended that the authors specify the model size of Qwen2.5-VL and the accuracy of its degradation descriptions at that scale. Since model inference relies on a VLM, its overhead should also be included in the complete pipeline. Comparisons with lightweight degradation recognition approaches such as CLIP-basd or DINO-based to verify whether the VLM’s advantages in degradation recognition accuracy or final restoration performance are sufficient to offset the additional computational cost.

Questions for authors
  1. What is the parameter size of the Qwen2.5-VL used in this paper (e.g., 3B, 7B, etc.)? Does it accurately recognize different degradation types and severity levels? My concern is that a smaller model without fine-tuning may fail to accurately recognize certain degradations. How can recognition accuracy and model size be balanced?

  2. The ablation study in Section 4.6 indeed reports results only on GoPro. I would like to know whether all ablation models use the same multi-task joint training setup?

  3. Additionally, is the random expert activation mechanism used only during training, or is it also enabled at inference time? My understanding is that it primarily serves as a training strategy to increase the diversity of expert participation through random activation. If it is also used during inference, please clarify how this randomness affects output stability and reproducibility, as well as the trade-off between the additional computational cost and performance gains.

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

The paper presents MoERestore for all-in-one image restoration under unknown and mixed degradations, aiming to combine general restoration ability, task adaptation, and texture recovery. It integrates shared experts, sparsely routed experts, and training-time random experts into a diffusion Transformer to allocate capacity according to the input. A vision-language model generates semantic degradation descriptions, while a pixel-space branch supplements detail information in the latent representation. Training combines diffusion, perceptual, high-frequency, and color-consistency losses. Using approximately one million paired images, the paper evaluates super-resolution, deblurring, dehazing, denoising, and mixed-degradation restoration, and uses a vision-language model to assess 60 real images.

Critical strengths
  1. Figure 1 and Sections 3.2–3.4 integrate semantic conditioning, sparse experts, and a pixel-space detail branch into a unified restoration pipeline. In Table 4, removing semantic conditioning, shared experts, random experts, or the high-frequency branch reduces performance on GoPro, indicating that these components contribute to the complete system.
  2. Table 1 reports an SSIM of 0.8341 on GoPro and PSNR values of 24.15 and 22.14 dB on the FoundIR blur-plus-noise and haze-plus-rain subsets, respectively. These results exceed the listed baselines.
Critical weaknesses
  1. Section 4.2 describes LucidFlux as the second-best method on DIV2K-Val, although Table 1 gives it the lowest PSNR/SSIM, and the proposed method has the worst LPIPS. Section 4.6 also attributes Table 4's largest performance drop to removing MoE, whereas it comes from removing the high-frequency branch. These ranking and attribution errors affect the assessment of performance margins, perceptual quality, and component contributions.
  2. The methodological contribution mainly combines components within established approaches. LoRA-IR already uses degradation-guided routing among low-rank experts, and FAPE-IR combines vision-language degradation planning, diffusion restoration, and frequency experts. The main changes in Sections 3.2–3.4 are the organization of shared, routed, and random experts and the pixel-space detail branch. However, what distinct patterns of sharing and conflict across tasks motivate this division remains unclear, leaving the contribution closer to an incremental combination.
Questions for authors
  1. Please correct the DIV2K-Val baseline ranking using Table 1 and explain why the proposed method has the worst LPIPS. After correcting the attribution of Table 4's largest performance drop, how does the interpretation of MoE and the high-frequency branch change?
  2. Section 3.2 emphasizes parameter efficiency. Including the semantic model, pixel-space branch, and diffusion sampling, what are the total and active parameter counts, latency, and GPU memory usage of the complete pipeline? Can these be compared at the same resolution?
  3. Compared with FAPE-IR, what patterns of sharing or conflict across tasks do the shared, routed, and random experts each address? What new modeling assumption or learning mechanism makes the explanation of reduced cross-task interference go beyond a generic mixture-of-experts configuration?
  4. Can Table 3's VLM scores be validated through independent human assessment or samples with reference images? Please report variability across repeated scoring and distinguish faithful recovery of textures and text from plausible newly generated details.
  5. Removing SDA in Table 4 only tests the presence of semantic conditioning. Do the degradation descriptions in Section 3.3 provide an independent benefit over simple task labels or generic prompts, particularly for mixed degradations?

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

Related papers

Discussion (0)

Sign in to comment.