Source-linked AI summary
BC-IHV: Conditioning the Color Space for Stable Rectified-Flow Low-Light Enhancement
Yi Ai, Zheng Chen, Yuanhao Cai, Yulun Zhang, Xiaokang Yang
TL;DR
LLIE must resolve ambiguous exposure without rewriting observable structure or chromatic content, while conventional intensity coordinates can impose unfavorable gradient behavior. The paper proposes SA-RF for structure-anchored rectified transport and BC-IHV for learnable intensity geometry. Across benchmarks and evaluations, the authors report consistent reconstruction and perceptual advantages, with controlled studies supporting both designs.
Problem
Generative LLIE can model ambiguous exposure but may modify observable geometry or color, while fixed color coordinates do not adapt inverse-gradient conditioning.
Method
SA-RF uses structure-aware conditioning and HybridAda, while BC-IHV provides a learnable analytically invertible Box–Cox polar intensity coordinate.
Results
The complete approach reports consistent reconstruction and perceptual advantages across three LOL benchmarks, blind evaluation, and cross-dataset transfer, with controlled ablations supporting its components.
Takeaways & Limitations
Jointly designing conditional transport and invertible color-space geometry supports generative low-light enhancement.
Takeaways & Limitations
The diagnostic comparison assumes only the intensity law changes while chromatic coordinates, velocity network, objective, and training schedule remain fixed.
Abstract
from arXiv · showhide
Low-light image enhancement (LLIE) must correct ambiguous exposure without overwriting structure already supported by the input. Generative transport can model exposure ambiguity; however, its flexibility may also alter observable geometry and chromatic content. Moreover, fixed invertible color coordinates are usually treated only as representations, although their inverse mappings reshape the RGB-domain gradients received by the enhancement network. To address these issues, we propose Structure-Anchored Rectified Flow (SA-RF), which maintains correspondence through separate chromaticity/intensity stems, a scale-matched condition pyramid, and HybridAda. HybridAda assigns location-specific retrieval to spatial cross-attention and global exposure modulation to pooled AdaLN. We further introduce BC-IHV, a learnable Box--Cox polar color space whose analytically invertible intensity mapping controls the inverse-gradient dynamic range through a single exponent. This allows the representation to balance dark-range expansion and gradient conditioning instead of adopting a fixed linear or logarithmic law. Experiments on three LOL benchmarks, blind image-quality evaluation, and cross-dataset tests demonstrate consistent reconstruction and perceptual advantages over the sota. Controlled studies further support the effectiveness of both the proposed framework and color representation.
Introduction
LLIE must recover uncertain exposure while preserving scene structure and chromatic evidence already visible in dark inputs. SA-RF and BC-IHV address this by constraining generative transport and learning intensity-dependent gradient conditioning.
- LLIE reconstructs normally exposed images from dark, noisy observations while scene layout, object boundaries, and much chromatic content are already observable.
- Generative transport models multiple plausible normal-light outputs but may alter geometry or color already supported by the input.
- HVI retains linear intensity, crowding dark values near zero, whereas logarithmic expansion can become substantially more sensitive under strong perceptual supervision.
- BC-IHV uses a learnable Box–Cox intensity law whose exponent controls inverse-gradient dynamic range, balancing dark-range expansion and gradient conditioning.
- SA-RF combines a scale-matched condition pyramid with HybridAda, separating location-specific structure retrieval from global exposure modulation.
- Experiments across three LOL benchmarks, blind evaluation, and cross-dataset transfer report strong reconstruction and perceptual performance, with controlled ablations supporting the design.
Related Work
Prior LLIE research includes deterministic restoration, generative transport, and color-space approaches. These lines of work motivate combining exposure modeling with structure-aware conditioning and learned intensity representation.
- Low-light restoration: Deterministic LLIE methods use Retinex decomposition, direct enhancement, or optimization-unrolled formulations with different image priors.
- Generative LLIE and flow matching: Generative LLIE models the one-to-many nature of exposure correction through normalizing flows, diffusion, and generative perceptual priors.
- Generative LLIE and flow matching: Rectified flow and flow matching replace iterative denoising targets with velocity regression along continuous transport paths.
- Color and intensity representations: Color representations separate illumination from chromatic structure, while prior approaches retain linear intensity, extend toward extreme darkness, or modify RGB brightness directly.
Problem Formulation
The formulation seeks a conditional distribution that transports ambiguous exposure while anchoring spatial and chromatic content visible in the low-light input. An invertible color transform determines both the latent geometry and the RGB-domain gradient path.
- The target is a conditional distribution p(x_rgb1 | y) that transports ambiguous exposure while anchoring spatial and chromatic content visible in y.
- An invertible color transform maps the low-light input and normally exposed target into a transformed coordinate system for transport.
- A velocity model transports Gaussian noise z ∼ N(0, I) to the target latent x1 conditioned on the transformed input.
- RGB-domain losses depend on the inverse transform, so Φ sets latent geometry and gradient flow while SA-RF controls transport and BC-IHV shapes RGB gradient scaling.
Method
The method combines Structure-Anchored Rectified Flow with BC-IHV to preserve observed structure during conditional transport while controlling RGB-domain gradient scaling through a learnable intensity coordinate.
- Structure-Anchored Rectified Flow: HybridAda separates local structure retrieval through spatial cross-attention from global exposure modulation through pooled AdaLN.Spatial tokens from c2 provide cross-attention keys and values, while global pooling conditions AdaLN.
- Structure-Anchored Rectified Flow: The conditional velocity field follows rectified-flow transport and produces the enhanced latent through ODE integration with 20 Euler steps.The transported state is processed by a conditional velocity network, and inference integrates dx/dt = uθ(x,t|c) from t = 0 to 1.
- Structure-Anchored Rectified Flow: SA-RF keeps observed multi-scale evidence available throughout transport using separate chromaticity/intensity stems and a scale-matched condition pyramid.The condition c = Φλ(y) forms {c0, c1, c2}, each entering the matching encoder and decoder resolution.
- Reconstruction Objective: Inverse-gradient scaling explains why comparable RGB residuals can produce different latent gradient magnitudes under logarithmic intensity mappings.The imbalance is isolated by κI, which measures the intensity-coordinate contribution rather than the entire network’s condition number.
- BC-IHV: BC-IHV uses an analytically invertible normalized Box–Cox intensity law whose exponent directly controls inverse-gradient dynamic range.Its closed-form inverse requires no learned inverse network, and λ is learned within [0.05, 1].
- BC-IHV: The representation retains HVI chromaticity while allowing the learned intensity law to interpolate between linear and logarithmic endpoints.BC-IHV adds a global scalar θλ relative to HVI; only the intensity mapping changes in matched comparisons.
- Reconstruction Objective: The training objective supervises both the color latent and reconstructed RGB image with fixed reconstruction, structural, edge, and perceptual losses.The perceptual weight remains fixed across representations to test sensitivity to inverse-gradient scaling.
Experiments
Experiments show strong reconstruction, perceptual, transfer, and efficiency results, while controlled studies attribute gains to both SA-RF and BC-IHV. The learned intensity law and architecture components consistently improve matched evaluations.
- Paired benchmark comparison: 28.37 dB PSNR on LOL-v1 exceeds InterLight by 0.50 dB, while LPIPS reaches 0.063 and SSIM matches the top reported precision.On LOL-v2 Real, the method instead leads LPIPS at 0.097; on Synthetic, it achieves best SSIM 0.941 and LPIPS 0.039 while remaining within 0.31 dB of highest PSNR.
- Paired benchmark comparison: SA-RF+BC-IHV improves CIDNet by 0.66, 0.31, and 1.41 dB on LOL-v1, LOL-v2 Real, and Synthetic, respectively.LPIPS falls by 0.016, 0.025, and 0.006, although LOL-v2 Real SSIM decreases by 0.009.
- Blind quality and cross-dataset transfer: On the three-LOL average, SA-RF+BC-IHV leads five blind metrics, while DICM/LIME tests rank it first across multiple metrics without in-domain fitting.It leads BRISQUE, MUSIQ, CLIP-IQA, MANIQA, and TOPIQ-NR on the LOL average, and transfers using LOL-v2 Real checkpoints.
- Qualitative comparison: Across five qualitative examples, SA-RF+BC-IHV obtains the highest displayed PSNR with clearer digits, boundaries, dark-region contrast, and fine textures.Cyan boxes identify shared enlarged regions across method outputs.
- Intensity law: 1.67 dB PSNR, 0.010 SSIM, and 0.034 LPIPS improvements over fixed HVI show that learned BC-IHV outperforms both fixed intensity endpoints under matched training.Against log-IHV, the gains are 1.22 dB, 0.022, and 0.066, respectively.
- Framework attribution: 1.41 dB cumulative PSNR improvement from HVI+CIDNet to BC-IHV+SA-RF indicates contributions from both the framework and representation.With HVI fixed, SA-RF raises PSNR by 0.94 dB; within SA-RF, BC-IHV adds 0.47 dB and gives the best SSIM.
- Architecture components: Removing any architecture component degrades all three metrics, with PSNR drops ranging from 0.52 to 2.62 dB.The ablation supports multi-scale conditioning and both HybridAda pathways.
- Intensity-law stress test: 24.55 dB versus 18.88 dB at the shared 430-epoch endpoint shows BC-IHV outperforming log-IHV under strong supervision.Learned exponents are non-endpoint values across LOL-v1, LOL-v2 Real, and LOL-v2 Synthetic, suggesting neither fixed endpoint is uniformly preferable.
Conclusion
SA-RF and BC-IHV jointly address structure preservation and color-space conditioning in generative LLIE. Experiments and ablations support their reconstruction and perceptual advantages across multiple evaluation settings.
- Conclusion: SA-RF keeps generative transport aligned with observed scene structure, while BC-IHV balances dark-range expansion and inverse-gradient conditioning.The framework uses multi-scale conditioning; the color representation uses a learnable Box–Cox coordinate.
- Conclusion: Experiments on paired benchmarks, blind evaluation, cross-dataset transfer, and controlled ablations demonstrate strong reconstruction and perceptual performance.
- Conclusion: Learned non-endpoint exponents support the proposed representation trade-off between dark-range expansion and gradient conditioning.