acceptodds
Under review as a conference paper at ICLR 2027

GraLoD: Graphics-Inspired Continuous Level-of-Detail Learning for Image Restoration

Abstract

The spatial support required for image restoration varies across degradation types, image regions, and reconstruction stages. However, most existing methods rely on predefined multi-scale hierarchies and aggregate features through fixed fusion or attention, leaving the representation scale itself largely determined by the network architecture. This limitation becomes more pronounced when a task-specific backbone is extended to heterogeneous degradations in all-in-one restoration. Inspired by level-of-detail (LOD) rendering in computer graphics, we propose GraLoD, a plug-and-play framework that treats restoration scale as a spatially varying and stage-dependent continuous variable. GraLoD reuses the native encoder hierarchy, aligns its multi-scale features into a shared LOD representation space, and predicts a stage-conditioned LOD field at each decoder stage. Each spatial location then continuously queries only two neighboring representation levels, enabling the effective restoration scale to adapt to both local image content and reconstruction progress. To prevent degenerate or arbitrary scale selection, we further introduce minimal-sufficient footprint calibration (MSFC) together with structure-aware regularization (SAR) to encourage restoration-effective and spatially coherent LOD assignments. GraLoD can be directly integrated into existing restoration backbones without redesigning their fundamental feature-processing blocks. Extensive experiments demonstrate consistent improvements in task-specific and all-in-one restoration.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image restoration, Level-of-Detail, Graphics-Inspired

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_oLvyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GraLoD, a graphics-inspired, plug-and-play framework for adaptive multi-scale image restoration. It addresses the limitation of predefined feature hierarchies by treating restoration scale as a spatially varying and decoder-stage-dependent continuous variable. Specifically, GraLoD aligns multi-scale encoder features into a shared representation space and predicts continuous Level-of-Detail coordinates to interpolate between adjacent feature levels. Minimal-Sufficient Footprint Calibration and Structure-Aware Regularization are further introduced to encourage restoration-effective and spatially coherent scale selection. Experiments across task-specific and all-in-one restoration demonstrate consistent improvements over multiple backbones, supported by component ablations and generalization evaluations.

Critical strengths
  1. GraLoD improves multiple restoration backbones across diverse degradation types. Notably, it increases the average PSNR of ALGNet from 30.60 to 31.70 dB and Perceive-IR from 31.93 to 32.14 dB in the all-in-one setting. Additional evaluations on unseen degradations further support its practical applicability and generalization potential.
  2. The continuous LOD formulation introduces an explicit adjacency constraint on scale selection, distinguishing it from unconstrained multi-scale attention. MSFC and SAR further regularize scale selection through reconstruction-based calibration and spatial consistency. The ablation studies provide useful empirical evidence for the contribution of these components.
Critical weaknesses
  1. The core LOD query is a piecewise-linear interpolation between adjacent encoder levels, mathematically equivalent to spatial scale attention with adjacent-level sparsity constraints. Prior work has explored adaptive receptive fields and spatially varying multi-scale fusion, including Selective Kernel Networks [1], ASFF [2], Pyramid Attention [3], and scale-adaptive deformable attention [4].Although Table 5 shows that GraLoD outperforms scale attention, the comparison does not sufficiently isolate the effects of continuous LOD from feature alignment, auxiliary supervision, and model capacity. Consequently, the specific advantage of the proposed LOD formulation over existing adaptive fusion mechanisms remains unclear.
  2. The paper interprets predicted LOD coordinates as the spatial support required for restoration. However, interpolating adjacent feature levels does not necessarily imply a corresponding change in the effective receptive field. Moreover, MSFC derives its targets from auxiliary probe errors, which may not reflect the optimal scale choices of the actual decoder.The absence of direct analyses relating LOD predictions to local degradation scales or effective receptive fields weakens this interpretation. Additionally, measured parameter overhead, FLOPs, inference latency, and memory consumption are not sufficiently reported, leaving the practical efficiency of GraLoD unclear.

[1] Selective Kernel Networks [2] Learning Spatial Fusion for Single-Shot Object Detection [3] Pyramid Attention Network for Image Restoration [4] A Universal Scale-Adaptive Deformable Transformer for Image Restoration across Diverse Artifacts

Questions for authors
  1. Could the authors compare GraLoD against standard scale attention and adjacent-level-constrained attention under matched feature alignment, auxiliary supervision, model capacity, and computational budgets? This would help establish whether the improvements arise specifically from continuous LOD modeling rather than additional architectural components.
  2. Could the authors quantitatively validate the relationship between predicted LOD coordinates and the effective receptive fields or spatial extents of local degradations? In particular, how well do MSFC's probe-based targets reflect the optimal scale choices of the actual decoder? Reporting parameter overhead, FLOPs, latency, and memory usage would also help assess the practical value of the proposed mechanism.
Reviewer_LAn2OpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

In this paper, the authors propose a plug-and-play framework GraLoD to improve the performance of existing image restoration models (of multi-scale encoder-decoder architecture). The authors say that their work is motivated by level-of-detail (LOD) rendering in computer graphics. The proposed pipeline is composed of three sub-networks, i.e., a multi-scale alignment module AsA_s, a feature-interpolation module GsG_s, and a decoder-injection adaptor HsH_s. The other contributions of this paper include two loss terms, minimal-sufficient footprint calibration (MSFC) to encourage the network to use the closest scales, and structure-aware regularization (SAR) to encourages similar LOD coordinates within different regions while allowing scale changes across structural boundaries.

Critical strengths
  1. Different from previous encoder-decoder restoration networks that depend on inflexible multi-scale feature fusion, this work utilizes neural networks to make predictions for layer selection and multi-level feature fusion. Their design is inspired by level of details (LOD) in graphics. The reviewer thinks that this motivation is very clear and the paper writing is sound. Ablate studies verify that, each of the proposed components, such as LOD, MSFC, SAR, bring notable improvement over the baseline model ALGNet according to Table 4. The motivation of this paper is very clear.

  2. The quantitative results show that the model equipped with the proposed methods outperforms the recent restoration works (published in 2024~2026) on a variety of tasks, such as, task specific image restoration (deraining, desnowing, dehazing, and deblurring), all-in-one IR (deraining, desnowing, dehazing, deblurring, and denoising), and zero-shot generalization evaluation. The reviewer appreciates the authors' hard works. The persuasive experimental results, such as 34.52 dB on the GoPro dataset, demonstrate the proposed architecture and loss design makes a new state-of-the-art in multiple restoration tasks without any additional training images and prior knowledge of large-scale visual data.

Critical weaknesses

The reviewer did not find any critical flaw of this paper in the initial review phase. Instead, the reviewer lists some minor issues as follows.

  1. The author provides ablation study results to examine each component in Table 4. However, only ALGNet is used as the baseline method, without clarifying which dataset is used. So, it is hard to tell whether or not the benefits are architecture- and domain-dependent or not.

  2. It is hard for the reviewer to find the visible improvement in visual comparisons (Figure 3 and Figure 4). The main paper does not include an analysis figure such as Figure 5 in appendix. Either Figure 3 or 4 could be placed in appendix while a polished Figure 5 could be moved to the main paper.

Questions for authors
  1. Do the results from Table 4 apply to the other networks, such as StarIR and Perceive-IR? Which dataset(s) were used to conduct the ablation studies?

  2. Could the authors provide a formalized Figure 5 and put it in the main paper?

  3. Do the authors find any clear visual improvements, such as the deblurred pictures with moving cars and car plates from the GoPro test set?

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

This paper proposes GraLoD, a plug-and-play framework that brings level-of-detail (LOD) selection into image restoration backbones. The method aligns the native encoder’s multi-scale features into a shared LOD space, predicts a stage-conditioned, spatially varying continuous LOD field per decoder stage, and queries two neighboring representation levels via continuous interpolation. To avoid degenerate scale selection, it introduces Minimal-Sufficient Footprint Calibration (MSFC) and Structure-Aware Regularization (SAR). Experiments cover CNN, Transformer, and Mamba backbones in task-specific, all-in-one, zero-shot, unseen-degradation, and unseen-severity settings.

Critical strengths
  • The paper introduces a principled, continuous LOD formulation for restoration that interpolates only between adjacent scales along an ordered axis, preserving scale order and continuity as opposed to generic attention over all scales. MSFC is a thoughtful calibration mechanism that biases toward the smallest sufficient footprint, mitigating collapse to coarse scales or arbitrary multi-scale gating. SAR exploits structural affinities for spatially coherent scale maps.
  • The plug-and-play claim is tested across heterogeneous backbones and settings: CNN (ALGNet, FSNet), Transformer (ECFNet), and Mamba-based (StarIR) backbones in task-specific (Table 1) and all-in-one (Table 2) settings, plus zero-shot generalization (Table 3), an unseen degradation type (Table 6), and unseen severities (Table 7). Direction of improvement is consistent across settings, which is what a plug-and-play claim requires.
Critical weaknesses

The paper lacks direct evidence for the learned LOD field. A central claim of the paper is that GraLoD learns a spatially adaptive and stage-dependent LOD field, while MSFC and SAR help prevent scale collapse and encourage spatially coherent scale selection. However, the paper does not directly examine whether the learned LOD field exhibits these properties. The current experiments mainly demonstrate improvements in PSNR, but do not show how LOD values vary across image regions, degradation types, or decoder stages. I suggest including LOD heatmaps and distribution statistics, particularly comparisons with and without MSFC and SAR. These results would help verify whether the proposed modules actually produce the intended scale-selection behavior, rather than simply improving reconstruction performance.

Questions for authors
  1. The paper lacks efficiency reporting, which is important for a module invoked across scales and stages. What is the parameter, FLOP, and runtime overhead introduced by GraLoD (Aℓ, Gs, Hs, resampling) across the evaluated backbones and input resolutions? Please provide a per-backbone breakdown and report training vs inference overhead.
  2. Could you visualize and analyze learned LOD fields across decoder stages and tasks (e.g., histograms of λ, spatial maps, etc.)? Do they align with intuition (coarser early, finer later) and vary by degradation?

Comments on the paper

1 comment
Area_Chair_9GT3OpenReview ↗

AC Reminder: Please Follow Up on the Authors’ Responses

Dear Reviewers,

Thank you for your time and thoughtful reviews. We are now about halfway through the discussion period, and the authors have posted responses to your questions and concerns.

If you have not yet done so, please read their responses and any relevant revisions, and leave a follow-up comment visible to the authors indicating which concerns have been addressed and which remain unresolved. Please update your review and rating if warranted. If your assessment remains unchanged, a brief explanation would still be helpful.

Thank you for your continued engagement in the discussion.

Best regards,

Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.