Low-Light Enhancement by Generating Only What Lighting Cannot Touch

A dark photograph mixes three things: the objects themselves (their colors and edges, called reflectance), the light falling on them (illumination), and sensor noise that grows as light drops. Most learned enhancers try to undo all three with one network. We propose a different split, borrowed from how quantum engineers protect fragile information. In a decoherence-free subspace, information is stored in quantities that the dominant noise cannot change. The textbook case is the relative phase of the states $|01\rangle$ and $|10\rangle$ under a phase kick shared by both qubits: the kick factors out as a global phase. Smooth illumination plays the same role for an image. In log intensity, a gain shared across a region adds a constant, and local log-contrast is unchanged. So we would train a velocity-parameterized flow model only on these illumination-invariant coordinates. Such a model learns to predict, at each moment, the direction a sample should move from noise toward data. Illumination would be recovered by a separate, deliberately simple low-dimensional estimator.

The second borrowed idea handles noise that is not shared. When independent energy loss hits a dual-rail qubit, the state falls into a detectably invalid configuration. The error becomes a known erasure, a flagged loss, rather than a silent corruption. Photon-starved pixels behave the same way. A pixel that records zero or one photon carries almost no information about reflectance. We propose to flag such pixels and hand them to the generative model as missing values to fill. We would not treat them as noisy measurements to be smoothed. The claim is that these two moves together remove the brightness-range problem from the generative model and stop sensor noise from leaking into reflectance. That leak is the central failure of classical Retinex enhancement.

Core Claim

Retinex theory writes an image as $I = R \odot L$, where $R$ is reflectance, $L$ is illumination, and $\odot$ is element-wise product. Structure-Revealing Low-Light Image Enhancement Via Robust Retinex Model adds an explicit noise term, $I = R \odot L + N$, as an escape valve so that corruptions are not forced into $R$ or $L$. This paper motivates our work, because it isolates the failure we target. It also constrains our design, because it models all noise as a dense additive map. We argue that the escape valve is the wrong shape for extreme low light. The valid information lies in two different places. Smooth illumination is a shared perturbation that a choice of coordinates can cancel exactly. Photon starvation is a per-pixel loss that is better flagged than estimated.

Concretely, let $\ell = \log(I + \delta)$ be the log image, where $\delta$ is a small offset for numerical safety. We define invariant coordinates

$$z = \ell - G_s * \ell,$$

where $G_s * \ell$ is a smoothing of $\ell$ at scale $s$. Any illumination roughly constant over scale $s$ cancels from $z$, just as the shared phase cancels from a dual-rail superposition.

We model each raw pixel count as $k \sim \mathrm{Poisson}(\lambda R L)$, where $\lambda$ is exposure. Pixels where $\lambda L$ falls below a threshold form an erasure mask $M$. The flow model generates clean invariant coordinates $z_0$ conditioned on the observed $z$ and on $M$. It is trained with the velocity target

$$v_t = \dot{\alpha}_t z_0 + \dot{\sigma}_t \epsilon,$$

where $z_t = \alpha_t z_0 + \sigma_t \epsilon$ interpolates between data $z_0$ and Gaussian noise $\epsilon$. Under a zero-terminal signal-to-noise schedule, $\alpha_T = 0$ at the final step, so training ends at pure noise.

The output image is reassembled from the generated $z_0$ and a separately estimated illumination. That illumination is a smooth, low-dimensional curve, so the generative model never has to produce absolute brightness.

Why the velocity objective matters here specifically

Common Diffusion Noise Schedules and Sample Steps are Flawed shows two things. First, standard schedules leak low-frequency signal into the final training step. Second, the resulting models cannot generate very dark or very bright images and drift toward medium gray. Their fix is a zero-terminal signal-to-noise schedule. That fix requires a velocity target, because noise prediction becomes ill-posed when the input is pure noise.

Low-light enhancement is exactly a task with extreme brightness ranges, so this brightness bias is a live risk for any diffusion enhancer working in image space. Our design attacks the same problem from a second direction: brightness is removed from what the model generates at all. SiT supplies the method. It shows that linear interpolants with velocity targets stay numerically stable over the whole time interval and give straighter paths that need fewer sampling steps. We reuse that recipe unchanged.

Because the two fixes overlap, the interesting scientific question is whether they are redundant or complementary. The program is built to answer it.

Research Program

We would build four variants of one conditional flow-model backbone, crossing two choices. The first choice is the target: noise prediction with a standard schedule, or velocity prediction with a zero-terminal schedule. The second choice is the space: generation in full image space, or generation in the invariant coordinates $z$. All four variants share the same architecture, data and training budget. On the best cell we add the erasure mask $M$ and compare it with the dense noise map of the robust Retinex model.

The illumination estimator is deliberately weak: a per-image, monotone tone curve fit on the smoothed log image. Any gain therefore comes from the generative part, not a stronger lighting network.

Quantum decoherence serves as an analysis lens, not a component. The invariance holds exactly only when illumination is constant at scale $s$. The analogy predicts that failures will concentrate where that "collective noise" assumption breaks, such as shadow boundaries and point light sources. This mirrors the symmetry brittleness reported for physical decoherence-free subspaces under spatially varying noise. We will test the prediction by measuring error as a function of local illumination gradient, and by sweeping $s$.

Sampling would use few steps. Shortcut distillation in velocity space, following Shortcutting Pre-trained Flow Matching Diffusion Models is Almost Free Lunch, is an optional final stage. We include it only if 3-step sampling is needed for parity in runtime.

Baselines

The naive reference point is classical Retinex applied without learning. The image is decomposed, illumination is brightened by a gamma curve, and the parts are recombined. This shows how much noise amplification the task contains.

The first learned baseline is Deep Retinex Decomposition for Low-Light Enhancement (Retinex-Net). It trains decomposition and enhancement networks with a structure-aware smoothness loss and applies BM3D denoising to reflectance. It is the canonical "denoise the reflectance afterward" approach, and the paired low-light dataset released with it is our main benchmark.

The strongest existing method in our file is the one-stage transformer Retinexformer. Its entry in the research file was cut off, and no reported result found on Lacuna for Retinexformer on that paired benchmark. We will therefore retrain it from released code on identical splits and report our own numbers rather than quote any.

The robust Retinex model with its additive noise map is run as an optimization baseline with tuned regularization weights. The most dangerous competitor is our own image-space velocity model with zero-terminal schedule. It isolates whether the invariant subspace adds anything beyond the existing diffusion fix.

Fairness is enforced by equal training data. Each learned baseline gets the same hyperparameter search budget, of a fixed number of trials on a held-out split. Diffusion variants are compared at matched sampling steps and wall-clock time. We count the proposal as beating Retinexformer only if it improves the perceptual metric without losing more than a small fixed margin in fidelity, on both the real benchmark and the darkest synthetic tier. The two ablations that isolate the key ingredients are invariant space versus image space, and erasure mask versus dense noise map.

Why not just…?

Why not just train an image-space flow model with velocity prediction and zero-terminal SNR? This is the real threat. Common Diffusion Noise Schedules and Sample Steps are Flawed already removes the gray bias, and a large enough model might learn illumination invariance implicitly. The 2×2 factorial settles this directly. If the image-space velocity cell matches the invariant-space cell in fidelity, brightness error and noise-in-reflectance, we will report that the subspace is unnecessary. The remaining contribution would then be the erasure mask alone. We expect a gap mainly at the darkest exposures, where image-space models must output large brightness changes and invariant-space models do not.

Why not just denoise first and then run Retinex? Retinex-Net effectively does this with BM3D on reflectance. A denoiser, however, must assign some value to a pixel holding one photon, and that value is a guess presented as a measurement. Our synthetic photon-count experiment reports error separately on erased and non-erased pixels. This tells us whether flagging-and-generating beats denoising specifically where information is absent.

Why not just keep the robust Retinex noise map but learn it? A learned dense map $N$ is the natural upgrade of Structure-Revealing Low-Light Image Enhancement Via Robust Retinex Model. It is a fair alternative, and it is our mask ablation. If the dense map performs as well, the erasure analogy adds nothing practical, and we will say so.

Proposed Experiments

We evaluate on two sources. The first is the paired low-light dataset from Retinex-Net, scored by PSNR, SSIM and LPIPS against references. The second is a synthetic benchmark made by rendering well-exposed images through a Poisson–Gaussian sensor model at exposures spanning about two orders of magnitude. In the synthetic set the true erasure mask and true illumination are known. That lets us measure reflectance error directly, along with how much noise power leaks into the recovered reflectance.

The decisive checks are three. The first is whether the brightness error of the noise-prediction model shrinks in invariant space relative to image space, which would show the subspace alone removes the gray bias. The second is whether masked generation beats the dense noise map specifically on erased pixels at the lowest exposures. The third is whether error concentrates at high local illumination gradients, as the broken-symmetry analysis predicts. If that third prediction fails, our account of why the method works is wrong even if the scores look good.