Standard Candles Inside the Clipped Plateau

When a camera pixel receives more light than it can hold, it records the same maximum value no matter how much brighter the scene really was. Image-restoration networks fill these blown-out regions with plausible content, and current benchmarks reward them for doing so. Because a clipped pixel carries no measurement, nobody can say whether a restored highlight is a recovered value or an invented one. Particle physicists face a similar problem with detectors they cannot open. Their answer is the standard candle: a reference source of precisely known output placed inside the instrument, so that its response can be checked against truth while it runs. We propose building standard candles whose true output is known inside the clipped region of a camera sensor. They let us measure, for the first time, how much of a de-clipped highlight is physically correct.

The candle is a light source whose output in several color channels is locked to fixed, known ratios and monitored by an independent photodiode. When one channel clips, the others still read correctly, and the known ratios give the true value of the clipped channel. A scene containing such candles therefore carries its own ground truth exactly where restoration methods are forced to guess. We use this to build a real-capture benchmark that separates recoverable highlights from unrecoverable ones. We then test a deliberately modest claim: deterministic recovery is possible up to a measurable "clip depth," and beyond that every current method, generative or not, is guessing.

Core Claim

Exposure-correction work treats saturation as a single hard wall. Model-Based Image Signal Processors via Learnable Dictionaries names it a "ceiling effect," and RAW Image Reconstruction Using a Self-Contained sRGB-JPEG Image with Only 64 KB Overhead argues that inverse mapping through clipped values is impossible. Generative models are then used to cross it: ECMamba defers extreme cases to generative priors, and CLIER in From Abyssal Darkness to Blinding Glare uses CLIP guidance with diffusion refinement. These papers motivate our work, because they locate the gap without measuring it.

We claim the wall is not uniform. Color Shift Estimation-and-Correction for Image Enhancement showed that color channels clip at different light levels. In the band between the first channel clipping and the last, some information survives. A ratio-locked candle makes that information exact. Let a candle emit with relative channel ratios $r_c$ (known by construction), let $g_c$ be the per-channel sensor gain, let $s$ be an unknown overall brightness, and let $V_{\max}$ be the clip level. The recorded value in channel $c$ is

$$y_c = \min\left(V_{\max},\; g_c\, s\, r_c\right).$$

Any single unclipped channel $c'$ fixes $s = y_{c'}/(g_{c'} r_{c'})$, and hence the true value of every clipped channel. Recovery is therefore determined exactly when $\min_c g_c s r_c < V_{\max}$, and undetermined when every channel clips.

The candle gives truth at both depths. That lets us test three things. The first is whether learned methods exploit surviving channels in the partial-clip band; on the evidence of CSEC's learned "pseudo-normal" color target, we expect them to do so only approximately. The second is whether their outputs in fully clipped regions are any more radiometrically accurate than a trivial prior. The third is whether a de-clipper that recovers where the physics allows and abstains elsewhere beats them on the metric that matters for scientific imaging.

The physics side supplies the design pattern and also a warning. The performance of SiPM-based gamma-ray detector (GRD) of GECAM-C anchors its energy scale in flight with intrinsic lines at 37.4 keV and 191 keV. Its high-energy, low-gain regime up to 6 MeV has no in-flight candle and relies on ground calibration. The internal tracers of Development of a ¹²⁷Xe calibration source for nEXO and the krypton lines of Measurement of the inhomogeneity of the KATRIN tritium source electric potential are designed so they do not drive the detector toward saturation. Standard candles have been kept out of the clipped regime on purpose. Our work puts them there, in imaging, where the cost of doing so is low.

Research Program

We will build candles from narrowband emitters driven at fixed current ratios. Each candle has a co-mounted photodiode that logs absolute output frame by frame, so the ratios $r_c$ are measured rather than assumed. We will measure each sensor's per-channel gains $g_c$, and its behavior near full-well capacity (the maximum charge a pixel holds), on a bench with an exposure ladder. This step checks the linearity assumption in the equation above rather than taking it for granted. It matters because real sensors can show a soft knee, blooming into neighboring pixels, and in-camera highlight reconstruction.

Candles will be placed in three kinds of captured RAW scene. In the first, candles light diffuse targets, which tests exposure correction of the kind studied by MSEC, LCDP and REED. In the second, candles are seen in reflection through glass, matching the setting of Removing Reflections from RAW Photos. In the third, candles act as night-time point lights through contaminated lenses, matching Flare7K.

Each static scene gets two independent truths. One is a tripod exposure ladder fused into a linear high-dynamic-range image. The other is the single-shot candle prediction from surviving channels and the photodiode log. Agreement between the two validates the candle. Disagreement measures how far sensor non-linearity departs from the clamp model, which is a result in its own right. Inputs at chosen clip depths come from the longer exposures, rendered through a fixed pipeline to sRGB for methods trained in that space.

The method we contribute is intentionally simple. It is a physically constrained de-clipper: it solves the equation above wherever at least one channel, or the unsaturated wings of a point candle's spread, constrain the brightness, and it outputs an explicit interval or abstains where nothing does. It is a reference instrument for interpreting the benchmark, not a competitor to large restoration networks on perceptual quality.

Baselines

The naive reference point is to leave clipped pixels at $V_{\max}$. We also include a "constant prior" that fills fully clipped regions with the training-set median highlight radiance. Any learned method that cannot beat the constant prior inside the plateau is producing appearance, not measurement.

The strongest existing competitors are the published exposure correctors, run with their released weights and also retrained under one shared budget on the same synthetic-clipped RAW training set. CSEC is the most direct competitor because it explicitly models channel-asynchronous clipping; it reports PSNR 23.627 and SSIM 0.855 on LCDP. ECMamba reports 23.76 dB PSNR on the ME dataset. CLIER is the generative representative; we found no reported result on Lacuna for its numerical REED scores. For flare scenes, the strongest baseline is Uformer trained on Flare7K, reported at PSNR 26.98 and SSIM 0.890 on the real-world test set. For the multi-view variant, the baseline is Learning Neural Exposure Fields for View Synthesis, which reports over 55% MSE reduction on HDRNeRF.

All of these figures are whole-image perceptual or fidelity scores. We found no reported result on Lacuna for radiometric error inside clipped regions, which is the quantity we introduce.

Two ablations isolate the candle's contribution. The first removes the known ratios and has the de-clipper estimate $r_c$ from image statistics alone. The second removes the photodiode log and relies on the ladder alone. "Beating" a baseline means lower log-radiance error inside the plateau in the partial-clip band, at matched coverage, and correct abstention in the full-clip band. Coverage here means the fraction of truths falling inside a stated interval.

Why not just…?

Why not just synthetically clip RAW images, for example from the MIT-Adobe FiveK raws behind MSEC? This is the real threat, because clamping linear RAW at a lower $V_{\max}$ gives free, exact truth. It fails only if real sensors deviate from a hard clamp near full-well capacity, through a soft knee, blooming, or crosstalk. We will test this directly. If method rankings and in-plateau errors on synthetic clipping match those on candle scenes within capture noise, we will report that synthetic clipping suffices, and the candle's remaining value is limited to single-shot, in-situ validation of dynamic scenes.

Why not just use bracketed exposures as ground truth, as REED does? Bracketing needs static scenes and alignment. The REED authors acknowledge ghosting from motion and intrusions. Bracketing also assumes linearity up to the clip level, which is the very assumption under test. We use bracketing as one of our two truths and report where it and the candle disagree.

Why not just fit point-spread wings, as astronomers do for saturated stars? For isolated point sources this works, and we include it inside our constrained de-clipper rather than competing with it. It does not apply to extended highlights, diffuse surfaces, or reflections, which are most of the clipped area in MSEC-, LCDP- and REED-like scenes.

Why not just add dual-gain hardware? GECAM-C shows that dual-channel readout extends range from 6 keV to 6 MeV. Even so, its upper channel has no in-situ check, and the extreme transient GRB 221009A still pushes the sensing medium itself toward its limits. Hardware moves the wall but does not validate what lies near it. Candles are what make that validation possible.

Why not just trust generative inpainting? The evidence in the file is qualitative: Flare7K reports semantic confusion between flare and real lights, and CLIER caps refinement at five iterations to avoid artifacts. Our experiment makes this quantitative. If CLIER beats the constant prior inside full-clip regions, generative priors carry real radiometric information there and our pessimism is wrong.

Proposed Experiments

The decisive check is a plot of in-plateau log-radiance error against clip depth, where clip depth is the true value divided by $V_{\max}$. The plot is split at the depth where the last channel clips. We predict three outcomes. The constrained de-clipper should stay near capture noise up to that split and abstain beyond it. CSEC should track it only partially in the partial-clip band. All methods should converge toward the constant prior beyond the split.

Chromaticity error in clipped regions, measured as perceptual color difference (ΔE), tests CSEC's specific color-shift mechanism. Interval coverage tests whether the abstaining method is honest about its uncertainty.

A ranking-agreement test compares candle truth with synthetic-clipped RAW truth and settles the main reviewer objection. A reflection-scene test asks whether RAW-domain separation recovers the transmission layer behind a partially clipped candle reflection.

Finally, an all-views-clipped test places a candle that is saturated in every view of a capture. Under the well-exposedness weighting of Neural Exposure Fields that region receives no supervision. We test whether the candle's single-view constraint restores it.