acceptodds
Under review as a conference paper at ICLR 2027

Degradation-aware Dynamic Adaptation for Unpaired Low-Light Image Enhancement

Abstract

Pretrained diffusion models provide powerful natural-image priors for low-light image enhancement (LLIE), offering a promising way to reduce dependence on paired training data. However, existing unpaired diffusion-based methods typically employ a fixed conditioning mechanism shared by all inputs. This is suboptimal for real-world LLIE, where different images may exhibit substantially different combinations of illumination deficiency, color distortion, and structural degradation. To address this limitation, we propose Degradation-aware Dynamic Adaptation (DDA), a one-step unpaired LLIE framework that adaptively exploits pretrained diffusion priors according to the degradation characteristics of each input. DDA contains two complementary modules. The Degradation-aware Prior Generation (DPG) module analyzes the illumination, color, and structural degradation of the input and produces an image-specific representation that identifies its restoration demands. The Degradation-Driven State-Glimpse (DDSG) adapter then translates these demands into layer-compatible dynamic updates for the image-conditioning pathway, allowing the diffusion prior to be injected differently for different low-light images. Extensive experiments on LOL-v2-real and LSRW, together with downstream classification, detection, and semantic segmentation evaluations, demonstrate that DDA consistently outperforms existing unsupervised LLIE methods. Code will be released upon acceptance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Low-light image enhancement, stable diffusion, unpaired learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Wy1JOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Degradation-aware Dynamic Adaptation (DDA), a one-step, unpaired low-light image enhancement (LLIE) method built on a pretrained diffusion image-translation backbone. Its Degradation-aware Prior Generation (DPG) module extracts illumination, color, and structure descriptors and selects an expert to produce an image-specific embedding. The Degradation-Driven State-Glimpse (DDSG) adapter combines this embedding with compact statistics of the base image-conditioning key/value projections to generate low-rank, sample-dependent updates. Training uses the CycleGAN-Turbo-style unpaired adversarial and cycle-consistency objectives. The paper reports improvements on LOL-v2-real and LSRW, downstream recognition benchmarks, and additional no-reference datasets.

Critical strengths
  1. The paper identifies the limitation of shared image-conditioning projections for heterogeneous low-light degradations and proposes a reasonably cohesive combination of explicit degradation priors, expert fusion, and low-rank dynamic projection updates.
  2. On LOL-v2-real, DDA reports 22.17 dB PSNR / 0.8201 SSIM / 0.1743 LPIPS, versus 21.32 / 0.8073 / 0.1928 for the closely related GEFU baseline. The submission also evaluates LSRW and three downstream tasks.
  3. The ablations test DPG and DDSG components, individual prior types and views, and MoLE/multiple-adapter alternatives. Appendix Table 8 provides paired significance tests against GEFU, and Table 10 includes no-reference evaluation on five additional datasets. These are valuable additions and should be acknowledged.
Critical weaknesses
  1. The novelty of DDA requires further clarification, as several of its key components have been explored in prior work. GEFU (ICCV 2025) adapts pretrained diffusion models for unsupervised LLIE; LLDiffusion (PR 2025) incorporates degradation representations into diffusion-based LLIE; MoR-DASR (AAAI 2026) employs degradation-aware low-rank expert routing; and HyperLoRA (EMNLP 2024) generates low-rank adapters using hypernetworks. Although DDA combines illumination/color/structure priors through top-1 expert fusion with projection-weight statistics to generate dynamic K/V updates, its methodological novelty beyond existing approaches remains insufficiently demonstrated. Despite the comparisons in Appendix Table 9, more targeted and capacity-matched comparisons are needed to clarify its distinctive contributions.
  2. The base K/V matrices are fixed during inference, so their encoded descriptors are fixed for each projection layer; they do not themselves change with the input. In this sense, the descriptors can serve as layer-specific conditioning signals rather than evidence that the generated update is intrinsically compatible with the matrix it modifies. Table 3 already includes constant-state and learned layer-ID controls, which is helpful; nevertheless, these comparisons do not directly establish why the chosen weight statistics are essential. Permuting the descriptors across layers or comparing with parameter-matched learned per-layer codes would provide a clearer causal test.
  3. Appendix Fig. 5 shows that the cosine distance between generated K/V updates grows when gamma, saturation, or noise perturbations increase for the same image. This demonstrates sensitivity to controlled changes, but not that the representation isolates degradation from image content or that the resulting adaptation is selectively beneficial for each degradation type. A generic image-conditioned hypernetwork might also produce similar trends. The paper would be more convincing with cross-content controls, a degradation-blind capacity-matched dynamic adapter, and an analysis linking the predicted adaptation to improvements under each degradation.
  4. The method introduces three prior encoders, expert routing, weight-state encoders, and dynamic low-rank generators. Table 2 provides task-dependent inference times, but not trainable parameter counts, FLOPs, peak GPU memory, or training cost for DDA and GEFU. In particular, ACDC runtime increases from 0.94 s (GEFU) to 1.15 s (DDA), approximately 22%, which deserves explicit discussion when describing the overhead as modest. Appendix Table 9 controls the backbone, training data, and insertion points for MoLE and multiple adapters, but the relative parameter budgets are not reported.
  5. DPG uses four hard top-1 experts, yet the paper does not report selection frequencies, expert collapse/balance statistics, or whether assignments correlate with the three degradation factors. Moreover, the stated target includes spatially non-uniform illumination, while the type-specific representations are globally pooled before routing. It would be useful to assess images with localized/compound degradations and provide failure cases. Appendix Table 10 also shows a decrease in CLIP-IQA on NPE (0.5396 to 0.5144), which should be discussed rather than only emphasizing mean gains.
Questions for authors

Q1. Could the authors explicitly position DDA against GEFU, LLDiffusion, MoR-DASR, and hypernetwork-generated LoRA methods such as HyperLoRA, distinguishing conceptual novelty from a new combination of existing components? In particular, what is uniquely enabled by jointly conditioning on degradation-prior embeddings and fixed K/V weight statistics, rather than by (i) degradation-conditioned expert selection or (ii) an equally sized input-conditioned hypernetwork without weight-state descriptors? A concise mechanism-level comparison and parameter-matched ablations on the same LLIE backbone would be valuable. Please distinguish the differences in task setting (e.g., MoR-DASR addresses SR and HyperLoRA addresses NLP) rather than presenting them as direct LLIE equivalents. Q2. Can you demonstrate that DDSG changes are specifically associated with degradation rather than general visual content? For example, keep degradation fixed while changing scene content, compare against a similarly sized image-conditioned hypernetwork without DPG, and report performance by dominant degradation category. Q3. Since the pretrained base projections are fixed, can you clarify what is uniquely gained by using the four weight statistics in Eqs. (8)–(9), beyond assigning each layer a distinct embedding? How would performance change if the weight-state vectors were shuffled between layers or replaced with equally sized learned per-layer vectors? Q4 Could you provide matched parameter counts, FLOPs, peak inference memory, and training/inference costs for GEFU, the complete DDA, and the Appendix Table 9 adapter alternatives, measured under the same input resolution and hardware? Please also clarify whether all Table 1 competitors were retrained or taken from prior reports. Q5. What fraction of inputs is routed to each of the four experts? Do the expert assignments correlate with the offline degradation proportions defined in Appendix Eqs. (29)–(32), and does the expert routing remain stable across cameras or localized illumination variations? Q6. AGLLDiff is discussed as a related pretrained-diffusion LLIE method (Section 2), but is absent from the quantitative comparisons. Could the authors explain the omission or provide a comparable result where the test protocol permits?

7 comments

W1 & Q1: Distinction from Existing Methods.

Response: We thank the reviewer for the constructive suggestion. We clarify the conceptual novelty of DDA and provide parameter-matched comparisons below.

(1) Novelty and Mechanism-Level Comparison. DDA's novelty lies in jointly conditioning dynamic K/V updates on degradation priors and target projection-weight states, rather than simply combining existing components. Specifically, (a) Unlike GEFU [6], which uses fixed image-conditioning K/V projections, DDA adapts them per input; unlike LLDiffusion [1], which learns degradation representations through joint generation and enhancement, DPG explicitly constructs and fuses multi-view degradation priors to drive projection adaptation. (b) Unlike degradation-conditioned expert routing (MoLE [2], MoR-DASR [3]), which selects or combines predefined LoRA updates, DDSG generates input-specific updates conditioned on both degradation and target projection states. (c) Unlike input-conditioned hypernetworks without weight-state descriptors (HyperLoRA [4]), DDSG explicitly incorporates fixed K/V weight statistics to tailor updates to the target projections. Neither alternative explicitly models this joint conditioning, although similar mechanisms could in principle be constructed by extending them. MoR-DASR (SR) and HyperLoRA (NLP) are included only for mechanism-level comparison, not as direct LLIE equivalents.

MethodTask / SupervisionDegradation InformationAdaptation MechanismUses Target-Weight State
GEFULLIE / UnpairedIllumination-aware promptsFixed K/V projectionsNo
LLDiffusionLLIE / PairedLearned degradation representationsDegradation-guided diffusionNo
MoR-DASRSR / PairedCLIP-based degradation scoresLoRA rank expert routingNo
HyperLoRANLP / Cross-taskTask representationsHypernetwork-generated LoRANo
DDA (Ours)LLIE / UnpairedMulti-view degradation priorsInput-specific low-rank K/V updatesYes

(2) Parameter-Matched Ablations. Full DDA outperforms all parameter-matched alternatives on the same LLIE backbone. Since the original MoLE and multiple-adapter baselines in Appendix Table 9 use only 31.41M and 107.16M trainable parameters, respectively, compared with DDA's 242.87M, we introduce three capacity-matched variants under identical training and evaluation settings on LOL-v2-real [7]. The comparisons isolate the contributions of update generation, target-weight conditioning, and degradation representation.

VariantModificationTrainable Params (M)PSNR ↑SSIM ↑
(i) DPG + LoRA routingDDSG → MoLE-style expert routing242.8421.090.7834
(ii) DPG + HypernetDDSG → HyperLoRA-style (cxc_x only, no weight state)242.8721.600.8104
(iii) Learned RGB + DDSGDPG → Learned RGB encoder242.8721.340.7980
Full DDA–242.8722.170.8201

These results support the benefit of jointly incorporating explicit degradation priors and target projection states, rather than relying solely on expert routing or input-conditioned update generation.

W3&Q2: Degradation versus image content

(1) Cross-content control. We acknowledge that our original analysis does not explicitly disentangle degradation from image content. However, applying identical synthetic perturbations across different scenes cannot guarantee equivalent degradation characteristics due to their inherent differences. We therefore conduct an alternative cross-scene analysis to examine whether a shared degradation factor induces consistent dynamic adaptation across different contents. Specifically, we apply the same gamma transformation x′=xγx'=x^\gamma to 100 normal-light reference images from LOL-v2-real and measure the mean pairwise cosine distance between their V-projection updates. For γ={1.7,2.2,3.0}\gamma=\{1.7,2.2,3.0\}, the distances decrease from 0.12062 to 0.07667 and 0.03767, respectively. This indicates that stronger common darkening leads to increasingly aligned update directions across different scenes, supporting the role of degradation in driving dynamic adaptation.

(2) Degradation-blind dynamic adapter. To assess the benefit of explicit degradation modeling over generic image-conditioned adaptation, we replace DPG and DDSG with a learned RGB encoder and an input-conditioned hypernetwork (RGB+Hypernet), respectively, while maintaining the same number of trainable parameters as DDA (242.87M). DDA outperforms this baseline on LOL-v2-real by 1.29 dB PSNR and 0.0378 SSIM, supporting the effectiveness of our degradation-aware design.

(3) Performance across degradation types. We further group LOL-v2-real test images by their dominant degradation following Eqs. (29)–(32). As shown below, DDA consistently outperforms both Learned RGB+Hypernet and GEFU across illumination-, color-, and structure-dominant subsets, achieving PSNR gains of 1.13/1.75/0.89 dB over RGB+Hypernet and 1.22/0.24/1.19 dB over GEFU, respectively. These results demonstrate consistent enhancement benefits across different degradation types, without assuming strict independence from image content.

DegradationNMethodPSNR ↑SSIM ↑
Illumination33DDA (Ours)22.060.8380
Learned RGB+Hypernet20.930.8052
GEFU20.840.8248
Color37DDA (Ours)22.120.7978
Learned RGB+Hypernet20.370.7458
GEFU21.880.7945
Structure30DDA (Ours)22.350.8278
Learned RGB+Hypernet21.460.8021
GEFU21.160.8037
Overall100DDA (Ours)22.170.8201
Learned RGB+Hypernet20.880.7823
GEFU21.320.8073

W2&Q3: Contribution of weight-state information

(1) Projection-specific guidance. We agree that the pretrained K/V projections remain fixed during inference, making their weight-state embeddings constant for each projection. However, these descriptors capture projection-specific weight characteristics rather than merely encoding layer identities. By combining them with the input-dependent degradation prior cxc_x, DDSG generates degradation-conditioned updates guided by the characteristics of each target projection.

(2) Weight-state shuffling. To verify the importance of the correspondence between weight-state descriptors and their target projections, we further conduct a Weight-state Shuffling experiment. Specifically, we randomly permute the weight-state embeddings across projection layers, keep the permutation fixed throughout training, and retrain the model from scratch under otherwise identical settings. This preserves model capacity and weight-state information while disrupting their correspondence with the target projections. As shown below, shuffling reduces PSNR/SSIM from 22.17/0.8201 to 21.66/0.8039, corresponding to decreases of 0.51 dB/0.0162. Importantly, although the shuffled embeddings can still serve as unique layer identifiers, they no longer describe the actual weight characteristics of their assigned target projections. Even when DDSG is allowed to learn from scratch under this fixed reassignment, the performance degradation persists. Together with the learned Layer-ID baseline, this suggests that correctly aligned projection statistics provide a more effective inductive bias than arbitrary layer-specific identifiers.

SettingPSNR ↑SSIM ↑
Constant Weight State21.620.7838
Learned Layer-ID Embedding21.490.7944
Weight-State Shuffling21.660.8039
Full DDA (Ours)22.170.8201
Reviewer_eifMOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Degradation-aware Dynamic Adaptation (DDA), a one-step unpaired low-light image enhancement framework that constructs various per-image illumination/color/structure priors as embedding and further produces input-specific low-rank updates to the image-conditioning pathway of a pretrained diffusion model. DDA attempts to address the limitation of sample-agnostic condition injection in existing diffusion-based LLIE methods, and extensive experiments on standard benchmarks and downstream vision tasks demonstrate that DDA achieves promising performance among unsupervised LLIE methods and verifies the effectiveness of degradation-aware dynamic conditioning on downstream tasks.

Critical strengths

1.The idea of introducing input-adaptive prior into diffusion process as well as cycle-consistency is well-motivated. 2.Comprehensive and extensive experimental validation are conducted to show the effectiveness of DDA on unsupervised LLIE and downstream vision tasks.

Critical weaknesses

1.Not really a weaknesses, but as the method is trained upon powerful pre-trained diffusion models. It would be interesting to see the generalization to various low-light images ( nosie levels, image sizes, distributions, etc.) In addition to the current two LOL-v2-real and LSRW, other representative paired or unpaired benchmarks are encourage to show the capability of dynamic adjustments (to name a few: UHD-LL, SICE,DICM,LIME,MEF,NPE,VV, etc.)

  1. Related to above point, robustness or failure boundary under extreme degradation are not quantitatively/qualitatively characterized, leaving the applicable scope of the method insufficiently defined.

3.The multi-view priors themselves remain a collection of classical, hand-crafted image statistics . While the subsequent dynamic fusion and weight-state conditioning are novel, the paper does not demonstrate that these particular hand-crafted views are necessary versus a learned or more compact feature extractor operating on the same raw input. Though ablations show that removing any single view hurts, yet they do not compare against an end-to-end learned degradation encoder of comparable capacity.

4.The cycle-consistency training under dynamic adaptation needs more discussion. Bidirectional cycle consistency is applied for unpaired training, yet it is not clear whether the degradation definition needs to be strictly symmetric or not, and if there are cases where dynamic adaptation in the reverse direction introduces unexpected variations that hurt cycle consistency.

Questions for authors

1.How does the method generalize to various low-light images ( nosie levels, image sizes, distributions, etc.) ,especially when handcrafted degradation priors may be corrupted by strong sensor noise. In addition to the current two LOL-v2-real and LSRW, other representative paired or unpaired benchmarks are encourage to show the capability of dynamic adjustments (to name a few: UHD-LL, SICE,DICM,LIME,MEF,NPE,VV, etc.) 2. Related to above point, robustness or failure boundary under extreme degradation are not quantitatively/qualitatively characterized, leaving the applicable scope of the method insufficiently defined. 3.The multi-view priors themselves remain a collection of classical, hand-crafted image statistics . While the subsequent dynamic fusion and weight-state conditioning are novel, the paper does not demonstrate that these particular hand-crafted views are necessary versus a learned or more compact feature extractor operating on the same raw input. Though ablations show that removing any single view hurts, yet they do not compare against an end-to-end learned degradation encoder of comparable capacity. 4.The cycle-consistency training under dynamic adaptation needs more discussion. Bidirectional cycle consistency is applied for unpaired training, yet it is not clear whether the degradation definition needs to be strictly symmetric or not, and if there are cases where dynamic adaptation in the reverse direction introduces unexpected variations that hurt cycle consistency.

4 comments

W1&Q1: Generalization Across Diverse Low-Light Conditions and Datasets

Response: We first clarify the protocol. DDA is trained only on the unpaired EnlightenGAN dataset [1] so every test set in the paper is unseen during training (Appendix C.1). This includes LOL-v2-real, the two LSRW subsets captured by different cameras (Huawei and Nikon), and the real night-time images of CODaN, DARK FACE, and ACDC used for the downstream tasks. The results below extend this evaluation.

(1) Real-world benchmarks (Appendix C.6, Table 10). DICM, LIME, MEF, NPE, and VV have already been evaluated in the appendix, and we will explicitly reference these results in the main text for better visibility. Without fine-tuning, DDA outperforms GEFU (Wang et al., 2025) in MUSIQ across all five datasets, with the average score increasing from 55.98 to 57.50. For CLIP-IQA, DDA achieves higher scores on four datasets, with a lower score only on NPE (0.5396 → 0.5144), while maintaining a comparable average score (0.4764 → 0.4794).

(2) Additional high-resolution evaluation. We further evaluate both pretrained models on the full UHD-LL [2] test set (150 images) without fine-tuning. To avoid GPU memory overflow at native 4K resolution, both models process images at 1024×576, and their outputs are upsampled to 4K for evaluation using the same procedure. Under this consistent evaluation protocol, DDA outperforms GEFU [3] by 2.44 dB in PSNR, 0.0278 in SSIM, demonstrating favorable generalization to unseen UHD-LL scenes.

MethodPSNR ↑SSIM ↑MUSIQ ↑CLIP-IQA ↑
GEFU18.900.802826.030.5427
DDA (Ours)21.340.830626.080.5700

(3) Sensor noise and prior corruption. Considering that noise often increases as illumination decreases in real-world scenarios, we jointly increase noise and reduce brightness to simulate realistic low-light conditions. Specifically, on the LOL-v2-real test set, we apply the gamma transform xγx^{\gamma} (γ>1\gamma > 1 darkens the image, as described in Appendix C.3), with γ\gamma increasing from 1.0 to 3.0, while adding Gaussian noise with σ\sigma increasing from 0 to 0.01. As shown below, model performance progressively declines with increasing noise and decreasing illumination. (Note: This table will be converted into a figure in the revised manuscript for clearer visualization.) As stated in Appendix D, spatially entangled noise and severe sensor-specific artifacts remain a limitation of predefined priors.

Gamma (γ)Noise (σ)GEFU PSNR ↑DDA PSNR ↑ΔPSNR ↑
1.0021.3222.17+0.85
1.20.0050020.3621.18+0.82
1.50.0062516.5117.72+1.21
2.00.0075013.4714.73+1.26
2.50.0087512.6713.82+1.15
3.00.0100012.4313.45+1.02

W2&Q2: Robustness and failure boundary under extreme degradation

Response: As discussed in W1 & Q1, model performance progressively deteriorates as scene illumination decreases and noise increases, indicating reduced robustness under extreme degradation. Nevertheless, DDA consistently outperforms GEFU across all tested degradation levels.

W3&Q3: Handcrafted Priors vs. Learned Degradation Encoder

Response: We replace DPG with an end-to-end trainable RGB degradation encoder of matched capacity (0.741735 M vs. 0.741725 M parameters), following the degradation encoder design of LLDiffusion [4], while keeping DDSG, the pretrained backbone, and training settings unchanged. The comparison is presented below.

MethodPSNR ↑SSIM ↑
Learned RGB + DDSG21.340.7980
Full DDA (Ours)22.170.8201
While alternative learned encoder architectures or task-specific optimization may achieve comparable performance, these results suggest that handcrafted degradation priors can provide useful inductive biases for degradation modeling.

W4&Q4: Cycle Consistency under Dynamic Adaptation

Response: DDA retains the bidirectional cycle-consistency framework of CycleGAN-Turbo (Parmar et al., 2024), while introducing degradation-aware dynamic adaptation. We clarify the following points:

(1) Dynamic Adaptation in Reverse Reconstruction. Strict symmetry of degradation conditions is not required. Both directions share the same translator, while dynamic weight updates are independently generated from their respective inputs and are not constrained to be inverse mappings. To investigate the effect of reverse dynamic adaptation, we disable dynamic weight updates only during reverse reconstruction, keeping the base adapter and all other settings unchanged. At the same training iteration, the full model achieves 22.17 dB PSNR and 0.8201 SSIM, compared with 21.57 dB and 0.8104 without reverse dynamic adaptation.

VariantPSNR ↑SSIM ↑Cycle L1 ↓Cycle LPIPS ↓
w/o reverse dynamic adaptation21.570.81040.02490.0346
Full DDA22.170.82010.05640.0580

** (2) Role of Cycle Consistency.** Cycle consistency serves as an auxiliary regularization constraint during unpaired training, rather than an objective requiring strict invertibility. Importantly, only the forward enhancement mapping is used at inference. Although reverse dynamic adaptation increases cycle-reconstruction errors (L1: 0.0249 → 0.0564; LPIPS: 0.0346 → 0.0580), it improves the final enhancement performance (+0.60 dB PSNR and +0.0097 SSIM). This suggests that better cycle reconstruction does not necessarily lead to better enhancement quality, as the primary objective is forward enhancement rather than reverse reconstruction accuracy. We will clarify this distinction and acknowledge the observed trade-off in the revised manuscript.

References

[1] Y. Jiang et al., "EnlightenGAN: Deep Light Enhancement Without Paired Supervision," IEEE Transactions on Image Processing, vol. 30, pp. 2340–2349, 2021. Paper

[2] C. Li et al., "Embedding Fourier for Ultra-High-Definition Low-Light Image Enhancement," International Conference on Learning Representations (ICLR), 2023. Paper

[3] S. Wang et al., "From Enhancement to Understanding: Build a Generalized Bridge for Low-light Vision via Semantically Consistent Unsupervised Fine-tuning," ICCV, pp. 13804–13814, 2025. Paper

[4] T. Wang et al., "LLDiffusion: Learning Degradation Representations in Diffusion Models for Low-Light Image Enhancement," Pattern Recognition, vol. 166, Art. 111628, 2025. Paper

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

This paper presents Degradation-aware Dynamic Adaptation (DDA) for unpaired low-light image enhancement using a pretrained one-step diffusion model. The method comprises two components: a Degradation-aware Prior Generation (DPG) module that extracts and adaptively combines illumination, color, and structure cues, and a Degradation-Driven State-Glimpse (DDSG) adapter that generates input-dependent low-rank updates to the image-conditioning projections. These updates are conditioned on both the degradation representation and descriptors of the base projection weights. The model is trained through bidirectional unpaired image translation and evaluated on enhancement benchmarks and downstream classification, detection, and segmentation tasks.

Critical strengths
  1. The method addresses a practically relevant problem through a coherent design. The variability of real low-light conditions makes adaptive use of pretrained generative priors a worthwhile direction. Combining explicit image-derived cues with compact projection updates provides a plausible approach to handling this variability. Retaining unpaired training and one-step inference further improves the practical appeal of the framework.
  2. The experiments provide meaningful evidence of the complete system’s effectiveness. Evaluation covers both enhancement quality and performance with fixed downstream models, helping assess whether the enhanced images remain useful for visual recognition. The component ablations, including the constant-state and learned layer-ID variants, provide additional support for the design choices. Taken together, the results suggest that the proposed framework offers useful empirical improvements beyond producing visually brighter images.
Critical weaknesses
  1. The benefit of dynamic projection updates is not sufficiently isolated. Section 3.3 motivates DDSG by describing fixed K/V projections as sample-agnostic. However, the attention response in Eq. (7) already depends on the input image tokens and UNet features. In Table 3, the IP-adapter baseline replaces both DPG and DDSG, changing the conditioning representation and adaptation mechanism simultaneously. The constant-state and layer-ID controls are useful for assessing the state-conditioning branch, but do not resolve this distinction. A comparison retaining the same DPG features while using a static conditioner with a comparable parameter budget would better establish the specific contribution of dynamic projection updates.
  2. The degradation analysis supports a narrower conclusion than the paper draws. Section 4.3 shows that update distances increase with changes in gamma, saturation, and noise while scene content is held fixed. This provides useful evidence of degradation sensitivity. However, a generic image-conditioned model could also respond to these perturbations, and the analysis does not establish that the generated updates provide appropriate corrections for the corresponding degradation. More direct evidence connecting the inferred prior to restoration quality would strengthen the interpretation that DPG identifies restoration demands. This matters because that interpretation is central to the proposed framework, beyond the observed performance of the complete model.
Questions for authors
  1. How does DDA compare with a static conditioning mechanism that uses the same DPG features and a comparable trainable-parameter budget?
  2. Are the base K/V projections frozen or jointly trained, and is the update-generating hypernetwork shared across attention layers?
  3. Does the “w/o DPG” variant exhibit similar update-distance trends under the controlled perturbations described in Section 4.3?
  4. For the same scene at different degradation levels, does using the corresponding DPG prior improve restoration compared with using a prior from another degradation level?
4 comments

W1 & Q1: Contribution of Dynamic Projection Adaptation

Response: We thank the reviewer for this important suggestion. We agree that the original IP-Adapter baseline [1] simultaneously changes the degradation representation and the adaptation mechanism, making it insufficient to isolate the contribution of dynamic projection updates. We clarify the distinction and provide additional controlled comparisons below.

(1) Dynamic projections versus static conditioning. We acknowledge that conventional attention remains input-dependent through its query and image-token features, even when the K/V projection weights are fixed. Our distinction concerns the projection mappings themselves, rather than whether the attention responses vary across inputs. In static conditioning, the same learned K/V projection matrices are applied to all images. In contrast, DDSG generates input-specific low-rank residuals conditioned on the degradation prior cxc_x and the target projection-weight state, allowing the K/V mappings themselves to adapt to different degradation conditions.

(2) Existing parameter-matched comparisons. To disentangle the contributions of degradation modeling and dynamic adaptation, we have conducted additional parameter-matched experiments, as also reported in our response to Reviewer R2. All variants retain the same pretrained backbone, training data, and optimization protocol. Specifically, replacing DDSG with MoLE-style LoRA routing [2, 3] while retaining DPG reduces PSNR from 22.17 to 21.09 dB. Replacing DDSG with a degradation-conditioned hypernetwork [4] without weight-state information yields 21.60 dB. These results indicate that the proposed projection-state-conditioned update generation provides benefits beyond alternative input-dependent adaptation mechanisms. However, we acknowledge that these comparisons do not fully isolate dynamic projections from static conditioning.

(3) Parameter-matched static-adapter comparison. To directly isolate the contribution of dynamic projection updates, we introduce an additional baseline, DPG + Static adapter, which retains the complete DPG module but replaces DDSG with a conventional static image adapter. Specifically, the degradation-prior embedding cxc_x is projected into the image-conditioning feature space and combined with the CLIP image token [5], while the subsequent K/V projection matrices remain shared across inputs without dynamic updates. We match the trainable parameter count to that of full DDA (242.87M), while keeping the pretrained backbone, training data, and optimization protocol unchanged. This controlled comparison allows both models to exploit the same degradation information, isolating the benefit of input-specific K/V projection updates over static conditioning.

MethodTrainable Params (M)PSNR ↑SSIM ↑
DPG + Static adapter (Param.-Matched)242.8721.110.7651
DPG + LoRA routing242.8421.090.7834
DPG + hypernet242.8721.600.8104
Full DDA (Ours)242.8722.170.8201

W2 & Q4: Degradation Sensitivity and Restoration Relevance

Response: We thank the reviewer for this insightful comment. We agree that the original degradation-aware analysis primarily demonstrates the sensitivity of dynamic updates to degradation variations, but does not directly establish whether these updates improve restoration under the corresponding degradation conditions. We clarify the existing evidence and introduce a controlled prior-matching experiment to examine the relationship between degradation representations and restoration quality.

(1) Degradation sensitivity. Our original analysis in Section 4.3 and Appendix C.3 evaluates the dynamic updates induced by controlled illumination, color, and structural perturbations while keeping the underlying scene content fixed. The mean cosine distances between projection updates increase with the degradation gap, reaching 0.51, 0.26, and 0.16 for illumination, color, and structure, respectively. These results demonstrate that DDSG responds systematically to degradation variations. However, we acknowledge that such sensitivity alone does not establish whether the induced updates are beneficial for restoration, since a generic image-conditioned model may also respond to these perturbations.

(2) Restoration benefits of degradation-aware conditioning. (a) Beyond update sensitivity, our ablation experiments provide evidence that explicitly constructed degradation priors contribute to restoration quality. Removing DPG while retaining DDSG reduces PSNR/SSIM from 22.17/0.8201 to 21.20/0.7967. Furthermore, replacing DPG with a parameter-matched learned RGB encoder while retaining DDSG yields 21.34/0.7980, indicating that explicit degradation representations provide useful inductive biases beyond generic image features.

(b) Performance across degradation types. We further group the 100 LOL-v2-real [6] test images by their dominant degradation types following Appendix Eqs. (29)–(32). We compare DDA with GEFU [7] and a parameter-matched Learned RGB+Hypernet baseline (242.87M trainable parameters), which replaces DPG with a learned RGB encoder and DDSG with an input-conditioned hypernetwork without explicit degradation priors or projection-state conditioning. As shown below, DDA consistently outperforms both baselines across illumination-, color-, and structure-dominant subsets, achieving PSNR gains of 1.13/1.75/0.89 dB over Learned RGB+Hypernet and 1.22/0.24/1.19 dB over GEFU, respectively.

DegradationNMethodPSNR ↑SSIM ↑
Illumination33DDA (Ours)22.060.8380
Learned RGB+Hypernet20.930.8052
GEFU20.840.8248
Color37DDA (Ours)22.120.7978
Learned RGB+Hypernet20.370.7458
GEFU21.880.7945
Structure30DDA (Ours)22.350.8278
Learned RGB+Hypernet21.460.8021
GEFU21.160.8037
Overall100DDA (Ours)22.170.8201
Learned RGB+Hypernet20.880.7823
GEFU21.320.8073

These results demonstrate that degradation-aware adaptation provides consistent restoration benefits across different dominant degradation conditions. Nevertheless, they do not directly establish whether the degradation prior extracted from a particular input is more effective than a mismatched prior from the same scene, which motivates the following controlled experiment.

(3) Matched vs. Mismatched Degradation Priors. To determine whether DPG priors provide restoration-relevant information rather than merely reflecting degradation changes, we conduct a controlled prior-swapping experiment on 100 LOL-v2-real test images with gamma transformations (γ∈{1.0,1.5,2.2}\gamma\in\{1.0,1.5,2.2\}). For each degraded input, we replace only its DPG embedding cxc_x with that extracted from the same scene at another degradation level, while keeping the input and all other settings unchanged.

Input GammaPrior γ=1.0Prior γ=1.5Prior γ=2.2
γ=1.022.17 / 0.820121.99 / 0.819322.02 / 0.8193
γ=1.517.97 / 0.731318.76 / 0.743419.44 / 0.7543
γ=2.214.48 / 0.567414.87 / 0.578715.26 / 0.5888

With identical inputs, changing only the DPG prior alters restoration quality. Matched priors achieve an average PSNR of 18.73 dB, outperforming mismatched priors (18.46 dB) by 0.27 dB. Compared with the original-image prior, matched priors improve PSNR by 0.79/0.78 dB at γ=1.5/2.2\gamma=1.5/2.2, respectively. Although the stronger prior performs best at γ=1.5\gamma=1.5, these controlled results indicate that DPG embeddings provide degradation-dependent information that directly influences restoration quality, rather than merely exhibiting degradation sensitivity.

Q2: K/V Projection Training and Hypernetwork Sharing

Response: We clarify the training and parameter-sharing strategy as follows.

(1) K/V projection training. The original K/V projections of the pretrained UNet remain frozen. In contrast, the newly introduced image-conditioning K/V projections (WipkW_{ip}^k and WipvW_{ip}^v) are jointly optimized with DPG and DDSG during training. DDSG generates input-specific low-rank residuals that are added to these base projections. During inference, the learned base projections remain fixed, while the generated residuals vary across inputs.

(2) Hypernetwork sharing. Each adapted cross-attention layer has an independent hypernetwork with the same architecture but separate parameters. Within each layer, the hypernetwork jointly generates the dynamic parameters for both K and V projections, which are then reshaped into their respective low-rank factors. Thus, hypernetwork parameters are not shared across attention layers.

We will clarify these implementation details in the revised manuscript.

Q3: Update-Distance Trends Without DPG

Response: Yes. We evaluate the w/o DPG variant, which replaces DPG with an RGB-to-MLP encoder while retaining DDSG, on 100 LOL-v2-real images under the same illumination, saturation, and noise perturbations. Both models show positive correlations between degradation gaps and update distances. Under Gaussian noise, the mean within-scene Spearman correlation for V-update distances is 0.914 for full DDA and 0.983 for w/o DPG, with positive correlations in all 100 scenes.

These findings are consistent with the original objective of our analysis: to verify that DDSG-generated dynamic updates respond systematically to controlled degradation variations, rather than to demonstrate that such sensitivity is exclusive to DPG. However, sensitivity alone does not guarantee effective restoration. The higher Spearman correlation of w/o DPG indicates that its updates are more monotonically tied to the perturbation magnitude, but this does not imply better restoration: a generic RGB encoder can respond strongly to degradation while lacking explicit degradation-type information. DPG’s multi-view priors provide type-specific cues (illumination/color/structure) that guide more appropriate corrections. The w/o DPG variant achieves lower PSNR/SSIM than full DDA (21.20/0.7967 vs. 22.17/0.8201), indicating that explicit degradation priors provide additional benefits for restoration quality.

References

[1] H. Ye, J. Zhang, S. Liu, X. Han, and W. Yang, “IP-Adapter: Text Compatible Image Prompt Adapter for Text-to-Image Diffusion Models,” arXiv:2308.06721, 2023. https://arxiv.org/abs/2308.06721

[2] X. Wu, S. Huang, and F. Wei, “Mixture of LoRA Experts,” ICLR, 2024. https://proceedings.iclr.cc/paper_files/paper/2024/hash/ce806d8b4bf38cd92d483e5a0490d983-Abstract-Conference.html

[3] E. J. Hu et al., “LoRA: Low-Rank Adaptation of Large Language Models,” ICLR, 2022. https://arxiv.org/abs/2106.09685

[4] D. Ha, A. Dai, and Q. V. Le, “HyperNetworks,” ICLR, 2017. https://research.google/pubs/hypernetworks-2/

[5] A. Radford et al., “Learning Transferable Visual Models From Natural Language Supervision,” ICML, pp. 8748–8763, 2021. https://proceedings.mlr.press/v139/radford21a.html

[6] W. Yang, W. Wang, H. Huang, S. Wang, and J. Liu, “Sparse Gradient Regularized Deep Retinex Network for Robust Low-Light Image Enhancement,” IEEE Transactions on Image Processing, vol. 30, pp. 2072–2086, 2021. https://doi.org/10.1109/TIP.2021.3050850

[7] S. Wang et al., “From Enhancement to Understanding: Build a Generalized Bridge for Low-light Vision via Semantically Consistent Unsupervised Fine-tuning,” ICCV, pp. 13804–13814, 2025. https://openaccess.thecvf.com/content/ICCV2025/html/Wang_From_Enhancement_to_Understanding_Build_a_Generalized_Bridge_for_Low-light_ICCV_2025_paper.html

Comments on the paper

1 comment

Dear Reviewers,

Thank you all for your valuable feedback. As the discussion period is approaching its end, we would greatly appreciate it if you could review our responses and let us know if any concerns remain.

We are happy to provide further clarifications and look forward to your feedback.

Thank you for your time and consideration!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.