acceptodds
Under review as a conference paper at ICLR 2027

Making Darkness Contagious: Prior-Guided Synthesis for Low-Light Object Detection

Abstract

Low-light object detection is constrained by scarce annotations, while preset synthesis rules can miss the degradation combinations found in real target images. We propose Contagious Degradation Synthesis (CoDeS), which learns reusable degradation “strains” from unlabeled low-light images and transfers them to labeled normal-light scenes. CoDeS represents target appearance with seven physically motivated descriptors and trains a conditional variational autoencoder (CVAE) to learn their joint variation conditioned on discovered illumination modes. Sampled configurations guide an image-formation-inspired online renderer that generates varied low-light views while preserving source boxes and categories. We further introduce a lightweight Low-Light Feature Adapter (LFA) and jointly train it with the SAM 3 detector on these views using inherited source annotations. Our approach requires no target training annotations or paired normal-light/low-light images, and synthesis is confined to training. Experiments on ExDark, NOD, and DARK FACE show significant improvements over pretrained SAM 3. Comparisons with alternative synthesis methods further show that CoDeS generates images that more closely match the local appearance and luminance statistics of real low-light data.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Low-light object detection, Domain adaptation, Synthetic data generation, Degradation modeling, Feature adaptation

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_vhZ1OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2(+1)
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper presents Contagious Degradation Synthesis (CoDeS), a low-light image synthesis framework for object detection. The central idea is to extract seven interpretable degradation descriptors from unlabeled real low-light images and learn their joint distribution using an illumination-conditioned conditional variational autoencoder (CVAE). The sampled degradation configurations are subsequently transferred to labeled normal-light images through an image-formation-inspired rendering pipeline, allowing existing object annotations to be reused for low-light detector adaptation. The authors further introduce a lightweight Low-Light Feature Adapter (LFA) to adapt SAM 3 features to low-light conditions. Experiments on ExDark, NOD, and DARK FACE demonstrate consistent improvements over direct SAM 3 inference.

Critical strengths
  1. meaningful problem formulation. The paper addresses an important limitation of existing low-light detection methods: the scarcity of annotated low-light images and the mismatch between manually designed degradation distributions and real imaging conditions.

  2. An interpretable and systematic synthesis framework. CoDeS incorporates multiple degradation factors, including illumination, noise, blur, and color bias, into a unified synthesis pipeline. The use of measurable appearance descriptors provides interpretable controls over the degradation process, although the effectiveness and generalization ability to out-of-domain scenarios need further discussion.

Critical weaknesses

1. Unfair comparisons with existing methods weaken the experimental support for the main contribution.

A major concern is the fairness of the comparisons in Table 2. The proposed method employs SAM 3, a large-scale pretrained vision model, while several competing methods are evaluated using substantially smaller detection architectures, such as YOLOv3. Consequently, the considerable performance improvements over existing methods may be largely attributable to differences in backbone capacity and pretraining rather than the proposed degradation synthesis strategy.

Notably, the improvements over direct SAM 3 inference are relatively modest (2.8–3.2 mAP50 points), compared with the much larger margins reported against conventional domain adaptation methods. Although the supplementary experiments demonstrate the applicability of CoDeS to additional detection architectures, they do not fully resolve this comparison issue.

The authors are encouraged to provide comparisons with representative competing methods under the same detection backbone, initialization, and training protocol. Without such controlled experiments, it remains difficult to determine whether CoDeS genuinely advances the state of the art in low-light domain adaptation.

2. Oversimplified appearance modeling and limited novelty in degradation prior learning.

The proposed degradation prior represents each low-light image using only seven image-level scalar descriptors. However, real-world low-light degradation is inherently spatially heterogeneous. Illumination, noise, color shifts, and blur can vary substantially across image regions and exhibit complex spatial dependencies. For example, different light sources may produce locally varying color casts, while sensor noise is closely related to local signal intensity.

Although the renderer introduces spatially varying illumination fields and signal-dependent noise, these local degradation patterns are not explicitly learned from the target images. Instead, the CVAE models only the joint distribution of global descriptors, while much of the spatial variation is introduced through predefined rendering rules. This raises concerns about whether the resulting synthesis can adequately reproduce the local degradation characteristics and spatial correlations observed in real low-light images. Besides, extracting a sequence of parameters to depict low-light condition is not novel, as ELVIS [CVPR'26] (ELVIS: Enhance Low-Light for Video Instance Segmentation in the Dark) has proposed a similar method.

Furthermore, the use of a conventional CVAE to model a low-dimensional parameter distribution offers limited methodological novelty. The authors should provide stronger evidence that this formulation is necessary and sufficiently expressive compared with simpler statistical models. Otherwise, the whole framework seems to be an incremental combination of existing techniques.

3. Insufficient interpretation and justification of the illumination modes.

The illumination modes play a central role in the proposed conditional degradation prior, but their interpretation and necessity are not sufficiently established. Section 3.2 defines these modes using K-means clustering based on brightness and shadow contrast. However, it remains unclear whether clustering in this two-dimensional space can meaningfully distinguish the diverse illumination conditions encountered in real-world environments.

Table 6 demonstrates that three illumination modes achieve the best detection performance across the evaluated datasets. Nevertheless, this empirical observation alone does not explain the physical meaning of the discovered modes or their stability across datasets. A more detailed characterization of the illumination clusters, including representative image examples and their associated degradation statistics, would help justify the proposed conditional modeling strategy.

4. Target-specific degradation fitting raises concerns about cross-domain generalization.

The proposed framework learns a separate degradation prior from unlabeled training images of each target dataset and subsequently synthesizes training samples that approximate the corresponding target appearance distribution. While this setting is valid for unsupervised domain adaptation, it introduces a strong dependence on the availability and representativeness of target-domain observations.

As a result, the current experiments primarily demonstrate target-specific adaptation rather than the transferability of the learned degradation representation. The method may be sensitive to changes in illumination conditions, camera characteristics, or noise distributions that are not adequately covered by the target images used for prior fitting.

Cross-dataset generalization experiments would be particularly valuable. For example, the degradation prior could be learned from one low-light dataset and then used to train a detector evaluated on another dataset without refitting the prior. Such experiments would clarify whether CoDeS learns reusable degradation characteristics or primarily matches the statistics of a particular target domain.

5. Limited improvements from CVAE-based sampling and insufficient statistical validation.

Table 5 shows that empirical sampling of degradation vectors from real target images already provides substantial performance improvements. Replacing empirical sampling with the proposed CVAE yields an additional gain of only 0.5–0.9 mAP50 points on ExDark, depending on whether the LFA is enabled.

Given that learning the joint degradation distribution is presented as a central contribution, these relatively small gains warrant further investigation. Comparisons with simpler density estimators, such as Gaussian mixture models, as well as multiple-seed evaluations, would help establish whether the learned CVAE prior offers a statistically reliable and meaningful advantage.

In addition, the current distributional visualization mainly focuses on brightness and shadow contrast. More direct evaluations of the full seven-dimensional distribution and the dependencies among degradation factors would strengthen the technical justification.

Questions for authors
  1. [The most important one] Can the authors provide a fair comparison with representative low-light domain adaptation methods using the same detector backbone, pretrained initialization, and training schedule?

  2. Given that the degradation prior consists of image-level scalar measurements, how can the method represent spatial dependencies among illumination, noise, color shifts, and blur? Can the authors provide quantitative evidence that the synthesized local degradation patterns match those observed in real low-light images?

  3. What is the physical interpretation of the three illumination modes in Table 6? Can representative examples and degradation statistics be provided to explain the distinctions between these modes?

  4. How well does a degradation prior learned from one target dataset generalize to another low-light dataset without refitting? Can the authors provide cross-dataset evaluation results?

  5. How does CVAE-based sampling compare with simpler alternatives, such as Gaussian mixture models? Are the relatively small improvements over empirical sampling consistent across multiple random seeds?

  6. The paper uses labeled target validation data for checkpoint and configuration selection. Can the authors clarify whether the competing methods follow the same model-selection protocol and provide results under a target-label-free selection strategy?

9 comments
AuthorsOpenReview ↗

Response to Reviewer vhZ1 (Part 1/3)

We thank the reviewer for the constructive comments and helpful suggestions!

Q1: [The most important one] Can the authors provide a fair comparison with representative low-light domain adaptation methods using the same detector backbone, pretrained initialization, and training schedule?

A1: We add an ExDark comparison by adapting AT, 2PCNet, and ISP-Teacher to SAM 3 while retaining their adaptation mechanisms. These baselines use the same official SAM 3 pretrained initialization, frozen image/text encoders, optimizer, and learning-rate schedule as CoDeS. No target training detection annotations are used. Test mAP50 results are reported below.

MethodmAP50 (%)
AT + SAM 379.8
2PCNet + SAM 378.4
ISP-Teacher + SAM 379.0
CoDeS + SAM 380.9

CoDeS exceeds AT, 2PCNet, and ISP-Teacher by 1.1, 2.5, and 1.9 percentage points, respectively, with the same SAM 3 backbone and pretrained initialization. We will include this comparison in the revised paper.

Q2: Given that the degradation prior consists of image-level scalar measurements, how can the method represent spatial dependencies among illumination, noise, color shifts, and blur? Can the authors provide quantitative evidence that the synthesized local degradation patterns match those observed in real low-light images?

A2: The CVAE learns the joint distribution of degradation descriptors measured from real low-light images, including how these factors vary together under different illumination conditions. Sampling from this learned prior guides synthesis toward the degradation characteristics of real low-light images, while the renderer realizes their spatial effects. The illumination scalar aa controls a smooth field Fsp(p)∝2aP(p)F_{\mathrm{sp}}(p)\propto 2^{aP(p)}, where P(p)P(p) is a smooth random field. Noise includes a local-luminance-dependent term and a row-correlated component. Chromatic gains and blur act on the spatially modulated image before exposure and noise are applied (Appendix A.3).

We evaluate local patterns on DARK FACE using 2,400 real images and 2,400 synthetic images per method. Alongside multiscale Lab-patch SWD, we compare spatial correlation curves of regional brightness, noise residuals, chromaticity, and edge smoothness across horizontal and vertical offsets. Spatial error below is the RMS difference from the real-image correlation curves.

Synthesis methodLab-SWD ×103\times10^3 ↓Spatial error (8×88\times8) ↓Spatial error (4×44\times4) ↓
Gamma45.990.10380.1408
Gamma+PG45.980.09970.1363
Crafting98.510.10230.1331
Similarity Min-Max71.850.10530.1374
DEN20.20‾\underline{20.20}0.08810.1162
ELVIS249.820.11030.1332
CoDeS12.220.0924‾\underline{0.0924}0.1264‾\underline{0.1264}

CoDeS reduces Lab-SWD by 39.5% relative to DEN and achieves the second-lowest spatial error at both grid sizes. The mean within-image Spearman correlation between local brightness and noise residuals is 0.8194 for CoDeS versus 0.8123 for real images on the 8×88\times8 grid; the corresponding 4×44\times4 values are 0.8204 and 0.7932. These results provide quantitative support for the local appearance, brightness-dependent noise, and regional spatial structure produced by CoDeS. We will include these analyses in the revised paper.

AuthorsOpenReview ↗

Response to Reviewer vhZ1 (Part 2/3)

Q3: What is the physical interpretation of the three illumination modes in Table 6? Can representative examples and degradation statistics be provided to explain the distinctions between these modes?

A3: The three modes describe illumination conditions through overall brightness bb and shadow contrast ss. Lower bb indicates darker appearance, while larger ss indicates a stronger separation between mid-tones and deep shadows. Taking ExDark as an example, with the labels used in Figure 5:

  • C1: strong shadows, with illuminated regions surrounded by deep shadows, such as scenes lit by windows or vehicle headlights.
  • C2: relatively brighter scenes with milder shadows, where more of the scene remains visible.
  • C3: overall darker scenes, such as dim interiors and nighttime environments.

We summarize their degradation statistics below as mean ± standard deviation.

DescriptorC1 (N=706N=706)C2 (N=1,332N=1,332)C3 (N=835N=835)
Brightness bb−3.372±0.893-3.372 \pm 0.893−3.098±0.683-3.098 \pm 0.683−5.386±0.827-5.386 \pm 0.827
Shadow contrast ss3.510±0.7523.510 \pm 0.7521.559±0.5881.559 \pm 0.5881.484±0.6301.484 \pm 0.630
Visible noise n×103n\times10^33.553±2.7103.553 \pm 2.7104.493±3.4674.493 \pm 3.4672.115±1.6922.115 \pm 1.692
Blur rr0.035±0.0260.035 \pm 0.0260.032±0.0220.032 \pm 0.0220.036±0.0240.036 \pm 0.024
Spatial illumination variation aa1.408±0.5541.408 \pm 0.5540.951±0.4130.951 \pm 0.4131.673±0.7241.673 \pm 0.724

Statistics use valid estimates; the valid counts for aa are 705, 1,332, and 819, respectively. C1 and C2 mainly differ in shadow contrast, whereas C2 and C3 mainly differ in overall brightness. These complementary conditions provide an interpretable structure for learning how degradation factors vary across illumination regimes. We will include these descriptions and statistics in the revised paper.

Q4: How well does a degradation prior learned from one target dataset generalize to another low-light dataset without refitting? Can the authors provide cross-dataset evaluation results?

A4: We evaluate ExDark→NOD transfer by freezing the ExDark degradation prior and the complete CoDeS model trained with it. We directly test this model on all NOD test images, without refitting the prior, fine-tuning the detector, or selecting checkpoints using NOD validation results.

Evaluation settingNOD mAP50 (%)
Pretrained SAM 360.6
CoDeS: ExDark→NOD61.9
CoDeS: NOD-specific adaptation63.4

The frozen ExDark model improves over pretrained SAM 3 by 1.3%, with a gap of 1.5% to NOD-specific adaptation. These results demonstrate that the CoDeS model trained with the ExDark prior transfers useful low-light detection capability to NOD without additional fitting or adaptation. We will include this cross-dataset evaluation in the revised paper.

AuthorsOpenReview ↗

Response to Reviewer vhZ1 (Part 3/3)

Q5: How does CVAE-based sampling compare with simpler alternatives, such as Gaussian mixture models? Are the relatively small improvements over empirical sampling consistent across multiple random seeds?

A5: We add a three-component, full-covariance GMM comparison on ExDark. The same-seed comparison below combines the original Table 5 Empirical/CVAE results with the supplementary GMM evaluations.

Sampling strategyWithout LFA: mAP50 (%)With LFA: mAP50 (%)
Empirical79.180.0
GMM78.779.9
CVAE79.680.9

We can find that CVAE achieves the highest mAP50. Its gains over GMM and empirical sampling are 0.9 and 0.5 percentage points without LFA, increasing to 1.0 and 0.9 percentage points with LFA. These results support the effectiveness of the learned illumination-conditioned prior and its complementary contribution alongside LFA.

We further report the original paper seed and two supplementary training seeds, all without LFA.

Sampling strategy20260621 (paper)8675309104729
Empirical79.177.978.6
CVAE79.678.479.2

We can see that CVAE is higher in these three displayed entries.

Q6: The paper uses labeled target validation data for checkpoint and configuration selection. Can the authors clarify whether the competing methods follow the same model-selection protocol and provide results under a target-label-free selection strategy?

A6: We reran IAT, YOLA, and FRBNet with SAM 3 while retaining their low-light modules. All methods use real ExDark training images without synthetic data. For these three baselines, we use checkpoints saved after a fixed number of training epochs. Their results are reported below.

MethodmAP50 (%)
IAT + SAM 384.9
YOLA + SAM 382.3
FRBNet + SAM 385.0
Ours86.7

Target training detection annotations are not used for prior fitting or detector optimization. For the ExDark ablation in Table 5, all trained variants are evaluated using checkpoints from the same fixed training epoch. The fixed checkpoint also coincides with the validation-best checkpoint in this ablation, so the fixed-epoch rule retains the reported results below. We will clarify this checkpoint-selection protocol in the revised paper.

Training strategyLFAmAP50 (%)
Pretrained SAM 3 (direct inference)✗77.7
Empirical sampling✗79.1
Empirical sampling✓80.0
CVAE generation✗79.6
CVAE generation (CoDeS)✓80.9
Reviewer_vh8gOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This work is motivated by reusing annotations of day-light images for low-light detection. It proposes CoDeS, which learns the distribution of seven appearance descriptors from unlabeled low-light target images, conditioned on modes ("strains") found by clustering brightness and shadow contrast. Descriptors sampled from this learned prior drive a renderer that maps the day-light image into a low-light one while preserving the bounding boxes and labels. The authors further propose an adapter (LFA), inserted between SAM 3's frozen image encoder and its detector and trained jointly with the detector on these darkened images, to improve detection in low-light scenarios. Experiments against several baselines are provided to support the method, along with an evaluation of the distributional similarity of darkened images with real-world low-light images.

Critical strengths
  1. By design, CoDeS adds little computational cost. Sampling a 7-dimensional descriptor vector from the learned prior is inexpensive, and the renderer consists of simple per-pixel operations. Training with these synthetic images therefore costs almost the same as standard training. This compares favorably with diffusion- or GAN-based unpaired image-to-image translation methods, which require training a large generative model. However, this efficiency comes at the cost of expressiveness.

  2. Descriptor estimation, reliability weighting, handling of invalid measurements, and rendering checks are well-documented.

Critical weaknesses
  1. The proposed method cannot ensure faithfulness to the physics of the underlying scene when degrading a natural image. For eg. shadow effects cannot be introduced arbitrarily in a scene as it is dependent on geometry and lighting conditions. Similarly, illumination variations (say) from the headlights of a car are confined to a specific area. Motion blur is not only dependent on the object motion but also on the depth of the scene. In fact, blur can also be space-variant such as due to camera shake. Faithfully capturing combinations of these effects is even more challenging. Merely transferring sample degradations to well-lit images without accounting for the underlying 3D geometry, lighting conditions, and object/camera motion is not technically correct.

  2. Synthesized images are restricted to what seven global descriptors and a fixed, content-agnostic renderer can express. Hence, the pushforward of the day-light images through this renderer may not faithfully represent the rich distribution of real low-light images. Unpaired image-to-image translation methods (e.g., CycleGAN, diffusion Schrodinger bridges) might have an edge here, as they learn such effects directly from target scenes without any hand-designed forward model restricting them, but they are not compared and are only briefly mentioned through DLDA.

  3. Several supervised baselines (IAT, FeatEnHancer, YOLA, and FRBNet) are YOLOv3-based results taken from prior work on ExDark and DARK FACE, whereas the proposed method builds on SAM 3. SAM 3 (direct inference) already outperforms all YOLOv3-based supervision baselines on ExDark, and all domain adaptation baselines on every dataset, without any training. The reported margins may therefore reflect the strength of the SAM 3 backbone rather than the proposed method.

  4. In Sec 4.1, what exactly do the authors mean by evaluation is not clear. The use of datasets and the evaluation protocol lack clarity. The figure captions are hardly of any help in the expt section; there is not even a clear statement on what the source images are. The detector is trained separately for each target dataset (ExDark, NOD, and DARK FACE) which is a limitation.

  5. The distributional metrics in Table 3 are computed on only 1,200 images, which makes them unreliable. The FID values (all above 200) are unusually high, which is not surprising as already reasoned in first and second points. No reference values are provided, such as the FID between two independent sets of real target images, so the small differences between methods are difficult to interpret.

  6. Low-light images are typically affected by Poisson noise. The authors have considered only visible noise which does not mimic natural low-light conditions.

  7. Is the original scene content preserved while synthesizing the degradations? There are no supporting experiments to validate this.

Questions for authors
  1. Since the strain label is a deterministic function of brightness and shadow contrast, which are already part of the descriptor vector, it carries no additional information, and sampling a strain by its frequency and then sampling within it recovers, in principle, the unconditional descriptor distribution (up to the confidence weighting of the mode frequencies). Moreover, the real contours in Figure 5 appear contiguous rather than forming three separated modes. What justifies the clustering stage, which adds complexity? If the motivation is the limited expressiveness of VAEs for multimodal distributions, why not use a generative model that represents multimodality directly (e.g., a Gaussian mixture model or a diffusion model) without a separate clustering stage? Pl report the silhouette scores, and clarify whether K was chosen by silhouette (Appendix A.2.2) or by validation mAP (Table 6).

  2. Since SAM 3 (direct inference) already outperforms the YOLOv3-based supervised baselines on ExDark in mAP50, the authors should re-run representative supervised baselines on the same SAM 3 backbone.

  3. Sec. 3.1 states that target detection labels are unnecessary, yet target validation labels are used for checkpoint and configuration selection. Could the authors report results without using any target labels, for example with a fixed number of training epochs or the final checkpoint? Also clarify whether the baselines followed the same selection protocol.

  4. Could the authors report FID and KID on a larger number of images, together with a reference value computed between two independent sets of real target images? Also report wrt CLIP-MMD.

2 comments
AuthorsOpenReview ↗

Response to Reviewer vh8g (Part 1/2)

Thank you very much for your valuable suggestions! We have conducted some experiments as per your request.

For question 1:

(1) Clustering motivation

Our motivation is to better fit the heterogeneous degradation distribution of the target dataset. We observe characteristic illumination regimes, including darker scenes, relatively brighter scenes with smaller shadow gaps, and scenes with pronounced shadows. Clustering makes these regimes explicit, allowing a shared CVAE to learn how all seven degradation factors vary jointly within each regime. The strain label provides a useful conditioning structure for finite-capacity learning, even though it is derived from the descriptors; the exact marginal decomposition does not imply identical learned approximations. To test this, we compare VAE (K=1K=1) and CVAE (K=3K=3) on a common ExDark 80/20 descriptor split, with matched latent/hidden dimensions and training settings. CVAE reduces mean holdout MMD$^2$ by 38.9%, supporting the benefit of conditioning for the learned approximation. Values are mean ± SD over three prior-fitting and sampling seeds.

MethodHoldout MMD⁡2×1000\operatorname{MMD}^{2}\times1000 ↓
VAE (K=1K=1)6.821±0.7716.821\pm0.771
CVAE (K=3K=3)4.169±0.6604.169\pm0.660

The contiguous contours in Figure 5 reflect continuous illumination variation across these regimes. Their distinction does not require empty gaps between contours: on ExDark, K=3K=3 achieves a weighted silhouette score of 0.4016, supporting a meaningful partition. Table 6 further shows detection gains of 1.5%, 2.0%, and 0.5% over K=1K=1 on ExDark, NOD, and DARK FACE, respectively.

(2) Comparison with GMM.

According to your suggestion, we fit a three-component, full-covariance GMM to the same unlabeled ExDark training descriptor bank. EM uses three initializations, with the fit selected by target-training likelihood. GMM samples drive the same renderer with the same parameter bounds and acceptance thresholds. For a controlled comparison, all sampling strategies use the same clean source images and annotations, physical renderer, parameter constraints, SAM 3 initialization, LFA architecture, loss, optimizer, training schedule, and model-selection and evaluation protocols. Only the degradation-parameter sampling strategy differs. The Empirical and CVAE entries below are from Table 5.

Sampling strategyWithout LFA: mAP50 (%)With LFA: mAP50 (%)
Empirical79.180.0
GMM78.779.9
CVAE79.680.9

CVAE outperforms GMM both without and with LFA. With LFA, its 80.9% mAP50 exceeds GMM and empirical sampling by 1.0 and 0.9 percentage points, respectively. These results support the effectiveness of our conditional prior for detection. All strains share one CVAE; changing K=1K=1 to K=3K=3 adds only 512 parameters (1.4%). Diffusion could also serve as an alternative prior, but the shared CVAE offers direct sampling of all seven controls in one decoder pass.

(3) Selection of KK.

We select the final number of illumination modes according to downstream detection performance on the validation set, which favors K=3K=3 on all three datasets. The weighted silhouette score assesses geometric separation. We summarize both criteria below for K∈{2,3,4}K\in\{2,3,4\}.

DatasetKKWeighted silhouetteValidation AP50 (%)
ExDark20.349578.72
ExDark30.401679.49
ExDark40.349078.54
NOD20.406563.73
NOD30.351065.62
NOD40.369263.53
DARK FACE20.522546.49
DARK FACE30.457748.56
DARK FACE40.473046.33

Although silhouette favors K=2K=2 on NOD and DARK FACE, its differences from K=3K=3 are 0.0554 and 0.0648, respectively, while K=3K=3 improves validation AP50 by 1.89% and 2.07%. On ExDark, K=3K=3 achieves both the highest silhouette score and the best validation performance. These results support a partition that captures illumination variations useful for detection. We will revise Appendix A.2.2 to distinguish the silhouette diagnostic from the final selection by validation performance.

Q2: Since SAM 3 (direct inference) already outperforms the YOLOv3-based supervised baselines on ExDark in mAP50, the authors should re-run representative supervised baselines on the same SAM 3 backbone.

A2: We reran IAT, YOLA, and FRBNet with SAM 3 while retaining their low-light modules. All methods use real ExDark training images without synthetic data. For these three baselines, we use checkpoints saved after a fixed number of training epochs. Their results are reported below.

MethodmAP50 (%)
IAT + SAM 384.9
YOLA + SAM 382.3
FRBNet + SAM 385.0
Ours86.7
AuthorsOpenReview ↗

Response to Reviewer vh8g (Part 2/2)

Q3: Sec. 3.1 states that target detection labels are unnecessary, yet target validation labels are used for checkpoint and configuration selection. Could the authors report results without using any target labels, for example with a fixed number of training epochs or the final checkpoint? Also clarify whether the baselines followed the same selection protocol.

A3: Target training detection annotations are not used for prior fitting or detector optimization. For the ExDark ablation in Table 5, all trained variants are evaluated using checkpoints from the same fixed training epoch. The fixed checkpoint also coincides with the validation-best checkpoint in this ablation, so the fixed-epoch rule retains the reported results below.

Training strategyLFAmAP50 (%)
Pretrained SAM 3 (direct inference)✗77.7
Empirical sampling✗79.1
Empirical sampling✓80.0
CVAE generation✗79.6
CVAE generation (CoDeS)✓80.9

We will clarify this checkpoint-selection protocol in the revised paper.For the additional supervised baselines, please refer to our response to Q2 for their fixed-epoch checkpoint protocol.

Q4: Could the authors report FID and KID on a larger number of images, together with a reference value computed between two independent sets of real target images? Also report wrt CLIP-MMD.

A4: We double the DARK FACE evaluation from 1,200 to 2,400 images per set, using the same source images and real reference set across all synthesis methods. We also report a real–real reference using two disjoint sets of 1,200 real target images each, and add CLIP-MMD.

Synthesis methodFID ↓KID ×103\times10^3 ↓Lab-SWD ×103\times10^3 ↓W1(Y)W_1(Y) ↓CLIP-MMD ↓
Crafting235.28186.2998.510.173.48
Similarity Min-Max198.20‾\underline{198.20}135.6171.850.123.10
DEN214.31161.5820.20‾\underline{20.20}0.04‾\underline{0.04}2.93
ELVIS227.38171.37249.820.483.64
CoDeS197.83145.34‾\underline{145.34}12.220.022.98‾\underline{2.98}

ReferenceFID ↓KID ×103\times10^3 ↓Lab-SWD ×103\times10^3 ↓W1(Y)W_1(Y) ↓CLIP-MMD ↓
Real–real reference16.890.79——0.01

The original metrics retain the Table 3 evaluation protocol, including the KID subset estimator. CLIP-MMD uses CLIP ViT-L/14@336 features and reports RBF MMD$^2\times10^3$ (σ=10\sigma=10).

These results preserve the conclusion of Table 3: CoDeS achieves the lowest FID, Lab-SWD, and W1(Y)W_1(Y) among the compared synthesis methods, with its strongest advantages in local appearance and luminance. CoDeS also achieves the second-lowest CLIP-MMD. We will include these expanded results and the real–real reference in the revised paper.

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

The paper proposes Contagious Degradation Synthesis (CoDeS) for low-light object detection. CoDeS extracts seven appearance descriptors from unlabeled low-light target images, clusters illumination conditions, and trains a conditional VAE to model the joint variation of degradation factors. Samples from this prior are transferred to labeled normal-light images through an image-formation-inspired renderer, while a Low-Light Feature Adapter (LFA) is jointly trained with a SAM 3 detector. Experiments on ExDark, NOD, and DARK FACE report improvements over direct inference with pretrained SAM 3 and over several synthesis baselines. The paper also includes ablations on the sampling strategy, illumination modes, degradation dimensions, and detector architectures.

Critical strengths
  1. The paper connects target appearance measurement, conditional prior learning, annotation-preserving rendering, and detector adaptation into a single pipeline. The descriptor definitions, missing-value handling, and rendering checks are described in considerable detail.
  2. Experiments cover three low-light benchmarks and include ablations of sampling strategies, illumination modes, individual degradation dimensions, and multiple detector architectures. These results provide evidence that the proposed components can be useful in practice.
Critical weaknesses
  1. The paper does not provide a controlled clean-source-only baseline, an LFA-only baseline, or a same-backbone reimplementation of the strongest competing methods. Several supervised comparison results are reported with different detector architectures, while the domain-adaptation numbers may not use the same training protocol. Consequently, the reported gains cannot be confidently attributed to CoDeS.
  2. The renderer explicitly checks brightness and shadow contrast, but not noise, blur, chromatic biases, or spatial illumination variation. The appendix acknowledges that the accepted rendering distribution can differ from the learned target prior, but no acceptance rates, terminal descriptor errors, or distribution-shift measurements are reported. Together with the absence of multi-seed results, this substantially weakens the evidence for the main claim.
Questions for authors
  1. How is the final number of illumination modes selected for each dataset? Is (K=3) chosen using the weighted silhouette criterion, validation performance, or test performance?
  2. Can the authors report results over multiple random seeds, including mean, standard deviation or confidence intervals, especially for the small gains reported on DARK FACE?
  3. What are the terminal errors, acceptance rates, rendering failure rates, and accepted-distribution statistics for all seven descriptors, rather than only brightness and shadow contrast?
  4. What are the results of clean-source-only training, LFA-only training, and source training with standard photometric augmentation under the same SAM 3 and optimization protocol?
2 comments
AuthorsOpenReview ↗

Response to Reviewer VAzv (Part 1/2)

We thank the reviewer for the constructive comments and helpful suggestions!

Q1: How is the final number of illumination modes selected for each dataset? Is K=3K=3 chosen using the weighted silhouette criterion, validation performance, or test performance?

A1: We select the final number of illumination modes according to downstream detection performance on the validation set, which favors K=3K=3 on all three datasets. The weighted silhouette score assesses the geometric separation of the illumination descriptors. We summarize both criteria below for K∈{2,3,4}K\in\{2,3,4\}.

DatasetKKWeighted silhouetteValidation AP50 (%)
ExDark20.349578.72
ExDark30.401679.49
ExDark40.349078.54
NOD20.406563.73
NOD30.351065.62
NOD40.369263.53
DARK FACE20.522546.49
DARK FACE30.457748.56
DARK FACE40.473046.33

Although the silhouette criterion favors K=2K=2 on NOD and DARK FACE, its differences from K=3K=3 are 0.0554 and 0.0648, respectively, while K=3K=3 improves validation performance by 1.89% and 2.07%. On ExDark, K=3K=3 achieves both the highest silhouette score and the best validation performance.These results suggest that the three-mode partition captures illumination variations relevant to detection in the evaluated low-light scenes. This supports our motivation for clustering, which aims to represent useful illumination regimes for conditional degradation modeling. We will clarify the distinction between silhouette analysis and the final selection using validation performance in Appendix A.2.2 later.

Q2: Can the authors report results over multiple random seeds, including mean, standard deviation or confidence intervals, especially for the small gains reported on DARK FACE?

A2: We evaluate our method on DARK FACE using five random seeds: {20260621,20260622,20260623,20260624,20260625}\{20260621, 20260622, 20260623, 20260624, 20260625\}. The test mAP50 results (%) are reported below.

2026062120260622202606232026062420260625
48.850.550.450.250.3

Across the five seeds, Mean ± SD is 50.0±0.7%50.0 \pm 0.7\%. Compared with the SAM3 baseline of 47.4%, this represents an average improvement of 2.6%. All five seeds outperform the baseline, and the average gain is substantially larger than the standard deviation of 0.7%. These results support a consistent and meaningful detection improvement on DARK FACE across random seeds.

AuthorsOpenReview ↗

Response to Reviewer VAzv (Part 2/2)

Q3: What are the terminal errors, acceptance rates, rendering failure rates, and accepted-distribution statistics for all seven descriptors, rather than only brightness and shadow contrast?

A3: We report a rendering audit on 1,024 source images from DARK FACE.

Terminal errors. The following table reports the MAE and 95th percentile (P95) of the absolute difference between the requested and terminal descriptors. Errors are calculated over valid estimates for each descriptor. Noise quantities are multiplied by 10310^3 in both tables.

DescriptorMAEP95Valid estimates
Brightness bb0.1100.4441,024
Shadow contrast ss0.2070.463990
Noise n×103n\times10^30.7631.2581,024
Blur rr0.0100.0301,024
Chromatic bias cRGc_{RG}0.1350.371998
Chromatic bias cBGc_{BG}0.1080.327998
Spatial illumination aa0.3860.933811

Acceptance rates. Both the first-attempt and final acceptance rates are 100.0% (1,024/1,024) under the existing terminal brightness/shadow and image-validity checks.

Rendering failure rate. The observed rendering failure rate is 0.0% (0/1,024) in this audit.

Accepted-distribution statistics. Below, we report the mean ± standard deviation of the accepted sampled parameters and the terminal-image descriptor estimates. Terminal statistics use the valid sample counts listed above.

DescriptorAccepted sampled parametersTerminal-image estimates
Brightness bb−5.615±1.043-5.615 \pm 1.043−5.663±1.105-5.663 \pm 1.105
Shadow contrast ss1.507±0.6161.507 \pm 0.6161.508±0.5711.508 \pm 0.571
Noise n×103n\times10^31.289±0.7041.289 \pm 0.7040.798±1.2930.798 \pm 1.293
Blur rr0.028±0.0050.028 \pm 0.0050.033±0.0130.033 \pm 0.013
Chromatic bias cRGc_{RG}0.104±0.1480.104 \pm 0.1480.006±0.0980.006 \pm 0.098
Chromatic bias cBGc_{BG}0.027±0.1310.027 \pm 0.131−0.030±0.128-0.030 \pm 0.128
Spatial illumination aa1.503±0.3911.503 \pm 0.3911.263±0.4831.263 \pm 0.483

All sampled configurations were accepted in this cohort. Accordingly, the sampled-to-accepted empirical distribution has zero Wasserstein distance for each descriptor and zero joint empirical MMD². The tables additionally quantify the differences between the accepted parameters and their realization in the final images.

Q4: What are the results of clean-source-only training, LFA-only training, and source training with standard photometric augmentation under the same SAM 3 and optimization protocol?

A4: We add the requested controls on ExDark using the same SAM 3 initialization, labeled normal-light source images, and optimization protocol. The expanded ablation results are summarized below.

Training strategyLFAmAP50 (%)
Pretrained SAM 3 (direct inference)✗77.7
Clean source✗78.3
Clean source + LFA (LFA-only)✓78.3
Source + standard photometric augmentation✗79.0
Source + standard photometric augmentation✓79.4
Source + Empirical Sampling✗79.1
Source + Empirical Sampling✓80.0
Source + CVAE Generation✗79.6
Source + CVAE Generation (CoDeS)✓80.9

Here, LFA-only denotes clean-source training with LFA and no learned degradation synthesis. Standard photometric augmentation uses ColorJitter with brightness/contrast/saturation strengths of 0.4 and hue strength of 0.1.

The full method achieves 80.9%, exceeding standard photometric augmentation by 1.9 percentage points and its LFA-equipped counterpart by 1.5 percentage points. Within the synthesis ablation, adding LFA to CVAE Generation improves mAP50 from 79.6% to 80.9%, while replacing Empirical Sampling with CVAE Generation under the same LFA setting improves it from 80.0% to 80.9%. The similar clean-source results (78.3%) may reflect limited low-light appearance variation in the training inputs. These comparisons support the complementary contributions of the learned degradation prior and LFA, beyond the improvements obtained through source training and standard photometric augmentation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.