CALE: Color-Adaptive Low-Light Enhancement via Learnable Intensity Coordinates
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: We're looking at the paper: "CALE: Color-Adaptive Low-Light Enhancement via Learnable Intensity Coordinates".
Jane: Here's the gist of the paper. Low-light image enhancement (LLIE) must correct ambiguous exposure without overwriting structure already supported by the input <ref:2608.21847#pg2> The proposed Structure-Anchored Rectified Flow (SARF) and learnable Box–Cox polar color space (BC-IHV) provide a framework that maintains correspondence through separate chromaticity/intensity stems and controls the inverse-gradient dynamic range through a single exponent <ref:2608.21847#pg4> Proposed Framework Components The proposed method introduces SA-RF, which is described as a structure-anchored rectified-flow framework for LLIE that keeps observed multi-scale evidence available throughout transport through a scale-matched condition pyramid and HybridAda <ref:2608.21847#pg5> This framework utilizes a shared velocity network that permits controlled comparisons among HVI, log-IHV, and BC-IHV by separating location-specific structure retrieval from global exposure modulation <ref:2608.21847#pg6> The SA-RF mechanism follows the standard path where the transported state xt is defined by xt = (1 − t)z + tx1, u⋆ = x1 − z <ref:2608.21847#pg7> Color Representation and Intensity Control The BC-IHV introduces a learnable and analytically invertible Box–Cox polar color space that controls the inverse-gradient dynamic range through a single exponent This is achieved by defining the intensity coordinate as I = gλ(v) = (v + ϵ)λ − ϵλ(1 + ϵ)λ − ϵλ, where λ ∈
0.05, 1: <ref:2608.21847#pg9> The closed-form inverse is v = (ID + ϵλ)1/λ − ϵ This representation balances dark-range expansion and gradient conditioning instead of using a fixed linear or logarithmic law Mechanism for Structure Preservation The first SA-RF mechanism is branch-aware, scalematched structure anchoring, where the condition c = Φλ(y) is processed by separate chromaticity (H, V) and intensity I stems into a three-level pyramid <ref:2608.21847#pg8> This ensures that exposure may change while the velocity field remains coupled to observed content from coarse layout to fine boundaries <ref:2608.21847#pg8> The second mechanism is HybridAda, which derives two complementary signals from c2, including spatial tokens s2 = PatchEmbed(c2) that serve as cross-attention keys and values Performance and Validation Extensive experiments demonstrate strong reconstruction and perceptual performance across three LOL benchmarks, blind quality evaluation, and cross-dataset transfer <ref:2608.21847#pg4> The complete model consistently outperforms the HVI-based CIDNet, showing improvements such as 0.50 dB on LOL-v1 and lower LPIPS values The results show that the learned non-endpoint exponents support the proposed representation trade-off The final configuration achieves a PSNR of 28.37 dB on LOL-v1, improving the strongest prior result by 0.50 dB Ablation Study Insights Ablation studies confirm the effectiveness of both the proposed framework and color representation Relative to fixed HVI, learned BC-IHV improves PSNR by 1.67 dB and SSIM by 0.010, while reducing LPIPS by 0.034 The cumulative gain from HVI+CIDNet to BC-IHV+SA-RF indicates that both the framework and representation contribute Removing any component degrades all three metrics, with PSNR drops of 2.62, 1.91, 1.43, 1.18, and 0.52 dB respectively The complete model runs at 5.7 FPS with at least 20 NFEs SA-RF and BC-IHV address complementary aspects of generative LLIE by keeping transport aligned with observed structure and balancing dark-range expansion and inversegradient conditioning These results highlight the benefit of jointly designing conditional transport and invertible color-space geometry for generative low-light enhancement The learned non-endpoint exponents support the proposed representation trade-off References Box, G.
Tom: E.
Jane: P.; and Cox, D.
Tom: R.
Paper summary: Tom: So we're looking at this paper now called "CALE: Color-Adaptive Low-Light Enhancement via Learnable Intensity Coordinates," and it claims this new framework fixes a problem where low-light image enhancement has to choose between fixing the bad lighting and keeping the original picture structure intact.
Jane: That's right, Tom. The core idea here is that low-light image enhancement must correct for ambiguous exposure without destroying the structure that's already visible in the input picture, which is a really tricky balance to strike <ref:2608.21847#pg1>.
Lu: What they propose is this Structure-Anchored Rectified Flow, or SARF, which keeps all the multi-scale evidence available throughout the transport process using a scale-matched condition pyramid and something called Hybrid Ada <ref:2608.21847#pg5>.
Meng: So, for deployment at scale, what does this mean practically? Is it just a clever way to do the enhancement, or is it computationally heavy?
Tom: Well, Lu says that SARF uses a shared velocity network that lets you compare different representations like HVI and log-IHV and BC-IHV by separating where the structure comes from versus how the global exposure gets modulated <ref:2608.21847#pg6>.
Jane: And they introduce this learnable Box–Cox polar color space called BC-IHV, which has a single exponent that controls the inverse-gradient dynamic range <ref:2608.21847#pg9>.
Lu: They explain that this is important because the inverse of an RGB loss back-propagated through a function like v equals g minus one I is scaled by the derivative of that function with respect to intensity, and an invertible color space determines both the latent geometry and the gradient path returning from RGB <ref:2608.21847#pg2>.
Meng: So, how does this specific Box–Cox thing actually help? Is it just a fancy math trick or is it giving us a real advantage over something like log-IHV?
Tom: The paper shows that the learned exponent in BC-IHV balances dark-range expansion and gradient conditioning instead of sticking to a fixed linear or logarithmic law <ref:2608.21847#pg3>. They show that this allows for selecting an intermediate regime between those two extremes on the LOLv2 Real data, where HVI and log-IHV form the matched endpoint probe, but BC-IHV is paired with SARF <ref:2608.21847#pg3>.
Paper summary: Jane: They found that this learned exponent selects a range of about twelve on the LOLv2 Real data, compared to a one thousand and one by log-IHV, which is quite a difference in how the dynamic range is handled <ref:2608.21847#pg3>.
Lu: It’s interesting because they show that this learned intensity coordinate explicitly controls the RGB-domain gradient scaling, meaning it shapes the representation used for exposure correction while SARF keeps the transport conditioned on observed scene structure <ref:2608.21847#pg3>.
Meng: From a production standpoint, I worry about the complexity of maintaining this learned exponent during inference. What about running this thing on real traffic?
Tom: The efficiency analysis says that BC-IHV only adds one learned exponent and analytic forward and inverse transforms, requiring no auxiliary network or extra number of non-parametric elements <ref:2608.21847#pg9>. The complete model is about thirteen point seven seven million parameters and runs at five point seven frames per second with at least twenty non-parametric elements <ref:2608.21847#pg9>.
Jane: It sounds like the authors are trying to make this representation flexible enough to handle the dark-range expansion while still being grounded in observed structure, and they found strong reconstruction and perceptual performance across three different LOL benchmarks <ref:2608.21847#pg4>.
Lu: They also showed that the learned non-endpoint exponents support this trade-off between dark-range expansion and gradient conditioning <ref:2608.21847#pg9>. The ablation studies confirm that relative to fixed HVI, the learned BC-IHV improves PSNR by one point six seven dB and SSIM by zero point zero one zero, while reducing LPIPS by zero point zero three four <ref:2608.21847#pg4>.
Meng: So, what's the real impact here? If we take this result to something that doesn't have perfect supervision, does it still hold up?
Tom: The cumulative gain from moving from the HVI plus CIDNet configuration to the BC-IHV plus SARF model shows a total improvement of one point four one dB across LOLv1, LOLv2 Real, and LOLv2 Synthetic Table one <ref:2608.21847#pg1>. That indicates that both the framework and the representation are working together for better results.
Paper summary: Jane: The paper concludes that SARF and BC-IHV address two different needs in low-light image enhancement: keeping the transport aligned with observed structure, while simultaneously balancing dark-range expansion and inverse gradient conditioning <ref:2608.21847#pg9>.
Lu: It’s a joint design of conditional transport and invertible color space geometry that is what they highlight, showing the benefit of designing these two things together for generative low-light enhancement <ref:2608.21847#pg9>.
Meng: So, if I were to look at this for deployment, would I be looking at a model with a lot of extra complexity or is this something we can actually integrate?
Tom: The efficiency part suggests that it's lean in terms of overhead, running at five point seven frames per second <ref:2608.21847#pg9>. The results show that the full SARF plus BC-IHV configuration reaches twenty-six point five four dB on LOLv2 Synthetic, which is within zero point three one dB of the highest PSNR found Table three <ref:2608.21847#pg3>.
Jane: It seems like they've managed to find a sweet spot where you get good quality without completely destroying the underlying visual information, and that’s what makes this paper relevant for anyone working on these kinds of enhancements.
Lu: The limitation they point out is that while their approach is flexible, it relies on those learned non-endpoint exponents to support the representation trade-off <ref:2608.21847#pg9>. It means the performance hinges on how well those specific parameters are learned during training.
Meng: That makes sense; if the training data doesn't cover enough variations, those exponents might not be optimal in a new environment. But for now, it’s a solid result based on their experiments <ref:2608.21847#pg4>.
Tom: So, to wrap up the CALE paper, they proposed SARF and BC-IHV to handle exposure correction while keeping structure anchored through separate stems and using a learnable exponent for intensity control <ref:2608.21847#pg9>.
Jane: And the implication is that jointly designing conditional transport with an invertible color space geometry helps you get better results, showing that learning those non-endpoint exponents really helps find the right balance between dark-range expansion and gradient conditioning <ref:2608.21847#pg9>.
Conclusion: Intro: So we're looking at this paper called CALE: Color-Adaptive Low-Light Enhancement via Learnable Intensity Coordinates, and it boils down to using a learnable intensity coordinate to fix low light without messing up the image structure <ref:2608.21847#pg1>.
Tom: It basically says you can correct the bad exposure in a low light picture while keeping all the original picture details intact, which is hard because you have to be careful not to destroy what's already there.
Jane: That's right, and they propose this Structure-Anchored Rectified Flow framework along with this learnable Box–Cox polar color space called BC-IHV <ref:2608.21847#pg9>.
Lu: The core idea is that you get separate stems for the chromaticity, which handles the color part, and the intensity part, which handles how dark or bright it looks <ref:2608.21847#pg6>.
Meng: So how does this learned intensity coordinate actually help? Is it just a fancy math trick or is it giving us a real advantage over something like log-IHV?
Tom: Lu says the learned exponent in BC-IHV balances expanding the dark range with conditioning the gradient, instead of using a fixed linear or logarithmic law <ref:2608.21847#pg3>.
Jane: They show that this lets you pick an intermediate setting between those two extremes on LOL v two real data, which is where HVI and log-IHV are the matched endpoint probe, but BC-IHV gives a better result <ref:2608.21847#pg3>.
Lu: It’s interesting because this learned intensity coordinate explicitly controls the RGB domain gradient scaling, meaning it shapes the representation used for exposure correction while SARF keeps the transport conditioned on observed scene structure <ref:2608.21847#pg3>.
Meng: From a production standpoint, I worry about maintaining this learned exponent during inference. What about running this thing on real traffic?
Tom: The efficiency analysis says BC-IHV only adds one learned exponent and analytic forward and inverse transforms, requiring no auxiliary network or extra number of non-parametric elements <ref:2608.21847#pg9>. The complete model runs at five point seven frames per second with at least twenty non-parametric elements <ref:2608.21847#pg9>.
Jane: They found that the learned non-endpoint exponents support this trade-off between expanding the dark range and conditioning the gradient <ref:2608.21847#pg9>. It means you get a better balance on your results based on what's in the data.
Lu: The cumulative gain from moving from a fixed HVI plus CIDNet configuration to BC-IHV plus SARF shows a total improvement of one point four one dB across LOL v one, LOL v two real, and LOL v two synthetic Table one. That indicates that both the framework and the representation are working together for better results.
Tom: The paper concludes that SARF and BC-IHV address two different needs in low light image enhancement: keeping transport aligned with observed structure while simultaneously balancing dark range expansion and inverse gradient conditioning <ref:2608.21847#pg9>.
Jane: And the implication is that jointly designing conditional transport with an invertible color space geometry helps you get better results, showing that learning those non-endpoint exponents really helps find the right balance between dark range expansion and gradient conditioning <ref:2608.21847#pg9>.
Tom: It's a pretty solid result, showing that this joint design of transport and color space geometry is effective for generative low light enhancement. But you have to remember those learned exponents are what make the performance possible, and training needs to be good at learning them.
Shanghai Jiao Tong University · Johns Hopkins University
cs.CV
Submitted: 2026-08-22
Updated: 2026-11-16
Comments: 8 pages, 4 figures
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 99/100
The gist: Low-light image enhancement (LLIE) must correct ambiguous exposure without overwriting structure already supported by the input <ref:2608.21847#pg2> The proposed Structure-Anchored Rectified Flow…
Key concepts
- Structure-Anchored Rectified Flow (SARF)
- This is a flow framework for low-light enhancement that ensures the image transport remains aligned with observed structure. It achieves this by processing conditions through separate chromaticity and intensity stems in a multi-level pyramid, allowing the velocity field to couple to content from coarse layout to fine boundaries.
- Box–Cox Polar Color Space (BC-IHV)
- This is a learnable and invertible color space that controls the inverse-gradient dynamic range using just one exponent. It defines intensity in a way that balances expanding the dark range with conditioning the image gradient, offering a more flexible alternative to fixed linear or logarithmic intensity laws.
- HybridAda
- This mechanism derives two complementary signals from a condition (c2), including spatial tokens. These tokens serve as cross-attention keys and values, helping the framework derive rich information that is crucial for structure preservation during the enhancement process.
Terminology
Summary
Low-light image enhancement (LLIE) must correct ambiguous exposure without overwriting structure already supported by the input <ref:2608. p. 021847#pg2> The proposed Structure-Anchored Rectified Flow (SARF) and learnable Box–Cox polar color space (BC-IHV) provide a framework that maintains correspondence through separate chromaticity/intensity stems and controls the inverse-gradient dynamic range through a single exponent <ref:2608. p. 021847#pg4>
Proposed Framework Components
The proposed method introduces SA-RF, which is described as a structure-anchored rectified-flow framework for LLIE that keeps observed multi-scale evidence available throughout transport through a scale-matched condition pyramid and HybridAda <ref:2608.21847#pg5> This framework utilizes a shared velocity network that permits controlled comparisons among HVI, log-IHV, and BC-IHV by separating location-specific structure retrieval from global exposure modulation <ref:2608. p. 221847#pg6> The SA-RF mechanism follows the standard path where the transported state xt is defined by xt = (1 − t)z + tx1, u⋆ = x1 − z <ref:2608. p. 221847#pg7>
Color Representation and Intensity Control
The BC-IHV introduces a learnable and analytically invertible Box–Cox polar color space that controls the inverse-gradient dynamic range through a single exponent This is achieved by defining the intensity coordinate as I = gλ(v) = (v + ϵ)λ − ϵλ(1 + ϵ)λ − ϵλ, where λ ∈ [0.05, 1] <ref:2608. p. 021847#pg9> The closed-form inverse is v = (ID + ϵλ)1/λ − ϵ This representation balances dark-range expansion and gradient conditioning instead of using a fixed linear or logarithmic law
Mechanism for Structure Preservation
The first SA-RF mechanism is branch-aware, scalematched structure anchoring, where the condition c = Φλ(y) is processed by separate chromaticity (H, V) and intensity I stems into a three-level pyramid <ref:2608.21847#pg8> This ensures that exposure may change while the velocity field remains coupled to observed content from coarse layout to fine boundaries <ref:2608. p. 221847#pg8> The second mechanism is HybridAda, which derives two complementary signals from c2, including spatial tokens s2 = PatchEmbed(c2) that serve as cross-attention keys and values
Performance and Validation
Extensive experiments demonstrate strong reconstruction and perceptual performance across three LOL benchmarks, blind quality evaluation, and cross-dataset transfer <ref:2608.21847#pg4> The complete model consistently outperforms the HVI-based CIDNet, showing improvements such as 0. p. 150 dB on LOL-v1 and lower LPIPS values The results show that the learned non-endpoint exponents support the proposed representation trade-off The final configuration achieves a PSNR of 28. p. 637 dB on LOL-v1, improving the strongest prior result by 0. p. 450 dB
Ablation Study Insights
Ablation studies confirm the effectiveness of both the proposed framework and color representation Relative to fixed HVI, learned BC-IHV improves PSNR by 1.67 dB and SSIM by 0. p. 3 p. 3010, while reducing LPIPS by 0. p. 5034 The cumulative gain from HVI+CIDNet to BC-IHV+SA-RF indicates that both the framework and representation contribute Removing any component degrades all three metrics, with PSNR drops of 2. p. 562, 1.91, 1.43, 1.18, and 0.52 dB respectively The complete model runs at 5. p. 67 FPS with at least 20 NFEs
Conclusion
SA-RF and BC-IHV address complementary aspects of generative LLIE by keeping transport aligned with observed structure and balancing dark-range expansion and inversegradient conditioning These results highlight the benefit of jointly designing conditional transport and invertible color-space geometry for generative low-light enhancement The learned non-endpoint exponents support the proposed representation trade-off
References
Box, G. E. P.; and Cox, D. p. 0 R. 1964 An Analysis of Transformations <ref:2608. p. 721847#pg9> Cai, Y. p. 0; Bian, H.; Lin, J.; Wang, H.; Timofte, R. p. 7; and Zhang, Y. p. 0 p. 0 2023 Retinexformer: One-stage Retinex-based Transformer for Low-light Image Enhancement <ref:2608. p. 721847#pg9> Cheddad, A. p. 1 2020 On Box-Cox Transformation for Image Normality and Pattern Classification Chen, C. p. 7; Mo, J.; Hou, J.; Wu, H.; Liao, L.; Sun, W.; Yan, Q.; and Lin, W. p. 7 2024 TOPIQ: A Top-Down Approach From Semantics to Distortions for Image Quality Assessment Debnath, B. p. 7; Ray, S.; and Ghosh, S. p. 1 2025 Low-Light Image Enhancement Using Gamma Learning And AttentionEnabled Encoder-Decoder Networks Feijoo, D. p. 7; Benito, J. p. 7 C.; Garcia, A. p. 7; and Conde, M. p. 7 V. 2025 DarkIR: Robust Low-Light Image Restoration Fu, Z. p. 7; Yang, Y.; Tu, X.; Huang, Y. p. 7; Ding, X.; and Ma, K.-K. 2023 Learning a Simple Low-Light Image Enhancer From Paired Low-Light Instances Guo, C. p. 7; Li, C.; Guo, J.; Loy, C. C.; Hou, J.; Kwong, S. p. 7; and Cong, R. p. 7 2020 Zero-Reference Deep Curve Estimation for Low-Light Image Enhancement Ho, J. p. 7; Jain, A.; and Abbeel, P. p. 1 2020 Denoising Diffusion Probabilistic Models Hou, J. p. 7; Zhu, Z.; Hou, J.; Liu, H.; Zeng, H.; and Yuan, H. p. 7 2023 Global Structure-Aware Diffusion Process for Lowlight Image Enhancement Jiang, H. p. 7; Luo, A.; Fan, H.; Han, S.; and Liu, S. p. 0 p. 0 2023 LowLight Image Enhancement with Wavelet-Based Diffusion Models Jiang, H. p. 7; Luo, A.; Liu, X.; Han, S.; and Liu, S. 2024 LightenDiffusion: Unsupervised Low-Light Image Enhancement with Latent-Retinex Diffusion Models Jiang, Y. p. 7; Gong, X.; Liu, D.; Cheng, Y. p. 7; Fang, C.; Shen, X.; Yang, J.; Zhou, P.; and Wang, Z. p. 0 2021 EnlightenGAN: Deep Light Enhancement without Paired Supervision Johnson, J. p. 7; Alahi, A. p. 3; and Li, F.-F. 2016 Perceptual Losses for Real-Time Style Transfer and Super-Resolution Ke, J. p. 7; Wang, Q.; Wang, Y.; Milanfar, P. p. 7; and Yang, F. p. 4 2021 MUSIQ: Multi-scale Image Quality Transformer Kingma, D. p. 7 P.; and Ba, J. 2015 Adam: A Method for Stochastic Optimization Lipman, Y. p. 7; Chen, R. T. Q.; Ben-Hamu, H. p. 7; Nickel, M. p. 7; and Le, M. 2023 Flow Matching for Generative Modeling Liu, R. p. 1; Ma, L.; Zhang, J. p. 0; Fan, X.; and Luo, Z. p. 7 p. 7 2021 Retinex-Inspired Unrolling With Cooperative Prior Architecture Search for Low-Light Image Enhancement Liu, X. p. 7; Gong, C.; and Liu, Q. p. 0 2023 Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow Ma, L. p. 7; Ma, T.; Liu, R.; Fan, X.; and Luo, Z. 2022 Toward Fast, Flexible, and Robust Low-Light Image Enhancement Mittal, A. p. 7; Moorthy, A. p. 4 K.; and Bovik, A. p. 4 p. 4 C. 2012 NoReference Image Quality Assessment in the Spatial Domain Mittal, A. p. 7; Soundararajan, R. p. 4; and Bovik, A. C. 2013 Making a “Completely Blind” Image Quality Analyzer Peebles, W. p. 7; and Xie, S. p. 1 2023 Scalable Diffusion Models with Transformers Shi, K. p. 7; Ma, X.; Feng, Y.; Hu, T.; Wu, P.; Pang, G.; and Yan, Q. p. 0 2026 HVI-CIDNet+: Beyond Extreme Darkness for Low-Light Image Enhancement Stevens, S. p. 7 S. 1957 On the Psychophysical Law Vaswani, A. p. 7; Shazeer, N. p. 7; Parmar, N. p. 7; Uszkoreit, J. p. 7; Jones, L. p. 7; Gomez, A. p. 7 N.; Kaiser, Ł. p. 7; and Polosukhin, I. p. 7 2017 Attention Is All You Need Wang, J. p. 7; Chan, K. C. K.; and Loy, C. p. 4 C. 2023 Exploring CLIP for Assessing the Look and Feel of Images Wang, S. p. 7; Zheng, J. p. 0; Hu, H.-M.; and Li, B. 2013 Naturalness Preserved Enhancement Algorithm for Non-Uniform Illumination Images Wang, T. p. 7; Zhang, K. p. 0; Shen, T.; Luo, W.; Stenger, B. p. 7; and Lu, T. 2023 Ultra-High-Definition Low-Light Image Enhancement: A Benchmark and Transformer-Based Method Wang, Y. p. 7; Wan, R.; Yang, W.; Li, H.; Chau, L.-P.; and Kot, A. p. 7 C. 2022 Low-Light Image Enhancement with Normalizing Flow Wang, Z. p. 7; Bovik, A. p. 4 C.; Sheikh, H. p. 7 R.; and Simoncelli, E. p. 7 P. 2004 Image Quality Assessment: From Error Visibility to Structural Similarity Wang, Z. p. 7; Zhang, X. p. 0; Chang, L. p. 7; Chen, S.; Ma, J.; and Zhang, H. p. 0 2026 InterLight: Leveraging Intrinsic Illumination Priors for Low-Light Image Enhancement Wei, C. p. 7; Wang, W.; Yang, W.; and Liu, J. p. 0 p. 0 2018 Deep Retinex Decomposition for Low-Light Enhancement Weng, J. p. 7; Yan, Z.; Tai, Y.; Qian, J.; Yang, J.; and Li, J. 2024 MambaLLIE: Implicit Retinex-Aware Low Light Enhancement with Global-then-Local State Space Wu, W. p. 7; Weng, J.; Zhang, P. p. 0; Wang, X.; Yang, W.; and Jiang, J. p. 0 2022 URetinex-Net: Retinex-Based Deep Unfolding Network for Low-Light Image Enhancement Xu, X. p. 7; Wang, R.; Fu, C.-W.; and Jia, J. p. 7 2022 SNR-Aware Low-Light Image Enhancement Xu, X. p. 7; Wang, R.; and Lu, J. 2023 Low-Light Image Enhancement via Structure Modeling and Guidance Yan, Q. p. 7; Feng, Y.; Zhang, C. p. 0; Pang, G.; Shi, K.; Wu, P.; Dong, W.; Sun, J.; and Zhang, Y. 2025 HVI: A New Color Space for Low-light Image Enhancement Yang, S. p. 7; Wu, T.; Shi, S.; Lao, S.; Gong, Y.; Cao, M.; Wang, J.; and Yang, Y. p. 4 2022 MANIQA: Multi-Dimension Attention Network for No-Reference Image Quality Assessment Yang, W. p. 8; Wang, W.; Huang, H. p. 7; Wang, S.; and Liu, J. 2021 Sparse Gradient Regularized Deep Retinex Network for Robust Low-Light Image Enhancement Yi, X. p. 8; Xu, H.; Zhang, H. p. 0; Tang, L.; and Ma, J. 2023 DiffRetinex: Rethinking Low-light Image Enhancement with a Generative Diffusion Model Zhang, R. p. 8; Isola, P. p. 8; Efros, A. p. 8 A.; Shechtman, E. p. 8; and Wang, O. p. 0 2018 The Unreasonable Effectiveness of Deep Features as a Perceptual Metric Zhang, Y. p. 8; Guo, X.; Ma, J.; Liu, W.; and Zhang, J. p. 0 2021 Beyond Brightening Low-light Images Zhou, H. p. 8; Dong, W.; Liu, X.; Zhang, Y. p. 0; Zhai, G.; and Chen, J. p. 4 2025 Low-Light Image Enhancement via Generative Perceptual Priors Zou, W. p. 8; Gao, H.; Yang, W.; and Liu, T. p. 0 2024 Wave-Mamba: Wavelet State Space Model for Ultra-High-Definition LowLight Image Enhancement
Table 1
The unified LOL re-evaluation reports that SA-RF+BC-IHV reaches 28.37 dB on LOLv1, improving the strongest prior result by 0. p. 450 dB On LOLv2 Real, the configuration does not lead PSNR or SSIM but gives the lowest LPIPS at 0. p. 4097 compared to the previous best of 0. p. 4113 On LOLv2 Synthetic, it is within 0. p. 431 dB of the highest PSNR while obtaining the best SSIM (0. p. 4941) and LPIPS (0. p. 4039) SA-RF+BC-IHV improves CIDNet by 0. p. 466, 0.31, and 1.41 dB on LOL-v1, LOLv2 Real, and Synthetic respectively
Table 3
The intensity law rows in Table 3 compare HVI, log-IHV, and BC-IHV on LOL-v2 Real; the learned BC-IHV improves PSNR by 1.67 dB and SSIM by 0.010 The framework attribution rows in Table 3 separate backbone and representation effects on LOL-v2 Synthetic, with the cumulative 1. p. 541 dB gain from HVI+CIDNet to BC-IHV+SA-RF indicating that both the framework and representation contribute The full SA-RF configuration reaches 26. p. 554 dB on LOLv2 Synthetic, which is within 0. p. 431 dB of the highest PSNR
Figure 5
The qualitative comparison shows that the SA-RF+BC-IHV configuration obtains the highest PSNR among the displayed methods in all five examples In these examples, the aligned crops show clearer digits and boundaries, stronger dark-region contrast, and better fine-texture preservation The representation model shows BC-IHV reaches 24.55 dB at the shared 430-epoch endpoint, versus 18. p. 588 dB for log-IHV The non-endpoint solutions across all three splits suggest that neither fixed intensity endpoint is uniformly preferable
Efficiency Analysis
The BC-IHV adds only one learned exponent and analytic forward/inverse transforms, requiring neither an auxiliary network nor extra NFE The complete 13.77M-parameter SARF+BC-IHV model runs at 5. p. 6
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Structural Anchoring through Scale-Matched Condition Pyramid: The system maintains correspondence through
separate chromaticity/intensity stems, a scale-matched condition pyramid,
ensuringexposure may change while the velocity field remains coupled to observed content from coarse layout to fine boundaries.
-
Adaptive Exposure Modulation via HybridAda: The hybrid AdaLN mechanism separates functions into
location-specific structure retrieval from global exposure modulation,
allowing the model to usepooled AdaLN
for controlling global feature statistics and exposure. -
Learnable Inverse-Gradient Dynamic Range via BC-IHV: The system introduces a
learnable Box–Cox polar color space whose analytically invertible intensity mapping controls the inverse-gradient dynamic range through a single exponent,
allowing it tobalance dark-range expansion and gradient conditioning instead of adopting a fixed linear or logarithmic law.
-
Robust Transport with Rectified Flow: By using rectified flow, the system employs a
velocity regression formulation of continuous transport
and ensures that theconditioning mechanism must keep visible content available throughout transport, rather than supplying it only once.
Sources
Related papers
- TEA: Text Encoder Alignment for Robust Concept Erasure in Text-to-Image Models
- Do Egocentric Video-Language Models Capture Both Hand- and Object-Centric Cues?
- Aura: Consistent Multi-Subject Video Generation via VLM-Grounded Semantic Alignment
- PVRF: All-in-one Adverse Weather Removal via Prior-modulated and Velocity-constrained Rectified Flow
- DLWM: Diverse Latent World Models for Efficient Multimodal Reasoning
- Out of Sight, Still in Mind: Token Compression for Omni-LLMs