Multi-Light Certificates as Verifiable Rewards for Intrinsic Image Decomposition

Splitting a photo into the color of the surfaces (albedo) and the light falling on them (shading) is underdetermined. A dark patch could be dark paint or a shadow. When you own several photos of the same static scene under different lights, though, the ambiguity partly goes away. Paint edges stay put and keep the same brightness ratio in every photo. Shadow and highlight edges move or change strength. This proposal turns that observation into an automatic grader. Given a stack of photos of one indoor scene, each taken under a different flash direction and at several exposures, the grader checks a candidate albedo map one image edge at a time. It confirms the edges the stack proves are paint, rejects edges the stack proves are lighting, and declines to judge everywhere else.

We use this grader as the reward in reinforcement learning with verifiable rewards (RLVR), the post-training recipe where a model is rewarded by a deterministic checker rather than a learned judge. The model being trained is a generative single-image decomposition model. It never sees the extra photos, and the grader uses them only during training. The closest existing work, ReasonX, already post-trains intrinsic models (Marigold, PRISM) with group-relative policy optimization (GRPO). Its reward, however, comes from a 4B-parameter vision-language model answering "do these two points have the same albedo?" on sparse point pairs. Our question is whether a physics-derived certificate can replace that judge and do better. The certificate is cheap, dense, can abstain, and has error rates we can measure.

Core Claim

A reward computed from multi-illumination evidence is a stronger and safer post-training signal for single-image albedo prediction than a neural perceptual judge, provided the reward abstains where the physics is ambiguous. The certificate rests on a classical fact. Write $g_k(p)$ for the log-image gradient at pixel $p$ under light $k$. If $p$ lies on a reflectance edge, $g_k(p)$ is nearly the same for every light $k$ under which $p$ is lit and unclipped. If $p$ lies on a shading edge, $g_k(p)$ varies across lights. The verifier labels each edge pixel as certified-reflectance, certified-shading, or abstain, using two quantities: the dispersion of $g_k(p)$ across lights and the magnitude of its median. The exposure brackets let the verifier exclude saturated and noise-floor pixels before it computes these statistics. That is where the cross-exposure part of the design matters.

A candidate albedo map $\hat A$ earns a reward equal to the F-score of its edges against the certified labels on non-abstained pixels:

$$r(\hat A) = F_1\big(\{p : |\nabla \log \hat A(p)| > \tau\},\ \mathcal{C}_{\text{refl}} ;\ \text{excluding } \mathcal{C}_{\text{abstain}}\big)$$

Here $\tau$ is an edge threshold, $\mathcal{C}_{\text{refl}}$ is the set of certified reflectance edges, and certified shading edges count as false positives when they appear in $\hat A$. We add one sign-and-ratio term. It checks that the predicted log-albedo step across each certified reflectance edge matches the median step measured in the stack.

The reward cannot be satisfied trivially. Copying the input image into albedo puts certified shading edges into $\hat A$ and is penalized. Predicting a flat albedo misses certified reflectance edges and is penalized too. Pure cross-exposure invariance, the property Siamese methods enforce, does admit such trivial solutions. It must be paired with a reconstruction constraint for exactly this reason.

Research Program

Training data. The training substrate is A Dataset of Multi-Illumination Images in the Wild. It contains 1,016 indoor scenes across 95 rooms, each captured under 25 bounce-flash directions with HDR brackets of at least 20 bits. That gives the verifier both the illumination axis and the exposure axis it needs.

The model being trained. Following ReasonX, the policy is a diffusion-based intrinsic model. A diffusion model is a generator that produces an output by gradually denoising random noise, so different random seeds give different albedo samples for the same photo. For each training photo $I_1$, GRPO draws a group of albedo samples. The verifier scores each sample using the other 24 lights, and the model is pushed toward samples that score above the group mean. A KL penalty to the base model keeps it close to its starting behavior. Because the verifier is a handful of image operations, a rollout costs a fraction of an MLLM call. This directly addresses the overhead ReasonX reports. We also skip groups whose rewards have zero variance, in the spirit of the rollout pruning in DEPO.

Calibrating the verifier. The second component treats the verifier as an imperfect instrument rather than ground truth. Cai et al. show that rule-based verifiers in math have large false-negative and false-positive rates, and that correcting policy gradients for estimated error rates helps. We can do something math verifiers cannot: measure our verifier's error rates directly. We render synthetic multi-light stacks with known albedo, in the style of the 450,000 rendered ShapeNet pairs of Ma et al.. On these stacks we sweep the number of lights $K$, the exposure noise, and the material type. This produces measured false-positive and false-negative curves for each label. We then apply forward correction with those measured rates instead of an estimated appeal judge.

Which constraints we draw on. Three bodies of evidence shape the design. The constraint comes from OpenIllumination: inverse rendering drops from about 35.77 PSNR on matte objects to about 22 on translucent or specular ones. That is why the verifier abstains on pixels whose gradients behave like neither paint nor smooth shading. The motivation comes from ReasonX's own admission that point-pair rewards "may overlook global structural consistency." The edge construction comes from the Retinex gradient-partition losses of Ma et al. and Retinex-Net. We reuse their logic as a test rather than as a training loss.

Baselines

The naive reference point is the base Marigold or PRISM intrinsic model with no post-training. Every other number is reported as a change from it, on the same evaluation code. The file gives no absolute IIW WHDR for these base models; no reported result was found on Lacuna for that figure, so we will measure it ourselves.

MLLM-judge GRPO (ReasonX). This is the most direct competitor, because it asks the same question with a neural judge in place of a physical verifier. ReasonX reports a 9–25% relative WHDR reduction on IIW over its base models. Beating it means a larger relative reduction from the same base model. We will reimplement its recipe: InternVL2.5-4B trained on synthetic point-pair comparisons, then GRPO on unlabeled real images. Both arms get the same group size, the same number of policy updates, and the same training scenes. ReasonX may additionally use any extra unlabeled images it wants, which gives it an advantage rather than us. Its wall-clock cost will be reported alongside ours.

CroCoDiLight. The strongest absolute number on IIW in the file is 15.4% WHDR, from CroCoDiLight. It trains on about 57,000 pairs, including the Multi-Illumination dataset, using a frozen cross-view-completion encoder. It is the reference for whether a multi-illumination-trained single-image model can be beaten at all. Matching 15.4% would be a strong result. Beating it would be decisive. Since the base models differ, we will report it as a cited number and not claim a controlled comparison.

InverseRenderNet. InverseRenderNet reaches 21.4% WHDR with a differentiable self-supervised rendering loss. It is the older self-supervised reference.

Ablations isolating the key ingredients. Two ablations come from our own pipeline. The first removes abstention, so the verifier labels every edge pixel by thresholded dispersion. This tests whether refusing to judge ambiguous pixels is what prevents degradation. The second restricts the verifier to two lights, matching the paired setting of Ma et al. This tests whether many lights are needed or a Siamese pair suffices.

Why not just…?

Why not just turn the same consistency into a differentiable loss and backpropagate? Siamese training with reconstruction and gradient-partition losses works for diffuse scenes (Ma et al., Retinex-Net). This is a real threat. Our verifier's abstention and median steps are non-differentiable, but a soft surrogate is easy to write. Backpropagating through a diffusion sampler is also expensive but feasible with few-step fine-tuning. We will run that arm with the identical soft certificate, data, and compute. If it matches GRPO, we will report that the gain lies in the certificate, not in reinforcement learning.

Why not just compute pseudo-label albedo from the stack and fine-tune on it with masks? Integrating the median log-gradients gives a dense reference albedo, so this alternative is the most serious one. Our reason to expect a difference is that integration spreads errors from ambiguous regions into confident ones. A ranking reward only needs to compare samples and never commits to a target. That argument is not proof. Masked pseudo-label distillation is therefore a full arm in the comparison, and if it wins, the RL framing should be dropped.

Why not just keep a neural judge but make it better? Cai et al. report false-positive rates of 35–68% for imperfect verifiers in math. Ackermann et al. show that policies drift toward sharp optima that exploit proxy rewards. A judge's error rate on albedo is hard to measure without the very ground truth it replaces, whereas ours is measured on synthetic stacks. The controlled ReasonX comparison will settle this empirically.

Why not just train on more synthetic data? ReasonX and InverseRenderNet both report weak transfer from synthetic to real images. Our synthetic data is used only to calibrate the verifier, not to supervise the model, so that domain gap does not enter the policy.

Proposed Experiments

Main comparison. The primary benchmark is WHDR on Intrinsic Images in the Wild, with no IIW labels used in training. We compare the base model, ReasonX-style GRPO, our GRPO, the differentiable-surrogate arm, the masked-pseudo-label arm, and both ablations. Each arm is run with three seeds. MIT Intrinsics serves as a secondary check where full ground-truth albedo exists. A held-out room split of the Multi-Illumination data gives a dense, in-domain score. That score is partly circular because it is graded by the verifier, so we report it only as a diagnostic.

Verifier calibration. This check comes before any RL. On synthetic stacks, the certified-reflectance false-positive rate must stay below the measured disagreement rate of the MLLM judge on the same edges. If it does not, the premise fails and we stop.

Reward-hacking check. Abstained regions carry no reward, so the policy could quietly corrupt them. We will bin IIW comparison points by whether our verifier would have abstained on similar content: specular highlights, cast shadows, colored interreflections. We then report WHDR separately for those bins. Improvement concentrated in certified regions with degradation in abstained ones would show hacking. Flat or improved abstained-region WHDR would show that the signal transfers.

Scaling check. The final analysis sweeps the number of lights $K$ used by the verifier. If the multi-light certificate is the active ingredient, WHDR should improve from two lights toward the full 25 and saturate. With few lights, dispersion estimates are noisy, so certification coverage should grow with $K$.