When a vendor sells a "relightable" 3D reconstruction (a digital twin of a product, a room, or a radio environment), the buyer has no way to check whether the shadows and color bleeding in the photos were explained by real geometry and light, or simply painted into the surface colors. Painting them in is called baking. A baked asset looks perfect from the training viewpoints and falls apart under new lighting. Vendors also have good reasons not to reveal the asset before payment, because it is their product. We propose a cryptographic certificate that settles this without revealing the asset. The vendor commits to the asset and then proves, in zero knowledge (a proof that a statement is true without revealing the data behind it), three things. The asset's light transport is self-consistent under a full multi-bounce lighting model. The asset reproduces the customer's photographs within a stated error. And it predicts how the object should look under a held-out lighting condition chosen by the buyer.
The central design choice is to prove that a solution is correct, not to replay the computation that found it. Inverse rendering (recovering geometry, materials, and lighting from photos) takes minutes to hours of stochastic gradient descent through a differentiable path tracer, a renderer that can pass error gradients back through simulated light paths. Arithmetizing that optimization into a proof circuit would be hopeless. Checking a candidate answer is much cheaper. In a radiosity formulation over small disk-shaped surface elements, global illumination reduces to a sparse linear fixed-point equation. Checking a claimed solution of that equation costs a single matrix-vector product plus range checks, which is the kind of structured linear algebra that existing zero-knowledge machine learning (zkML) systems already prove efficiently.
Core Claim
A certificate of physical consistency can be built from three ingredients. The first is a residual check on the radiosity equation, which binds the committed emission, reflectance, and geometry to a committed multi-bounce radiance solution. The second is a relighting check that renders the committed asset under a buyer-chosen environment light. The third is a set of physical sanity constraints: non-negative emission and reflectance at most one. Together these can detect baked or fabricated assets at prover costs comparable to proving a mid-sized vision network. The claim has two falsifiable parts. Binding the relighting check to the committed asset must be what defeats cheaters; an unbound relighting check or a direct-lighting-only check should not suffice. And fixed-point arithmetic must preserve relighting quality on real benchmarks.
Concretely, let $B$ be the vector of outgoing radiance per surface element, $E$ the emission, $\rho$ the per-element reflectance, and $F$ the sparse form-factor matrix, whose entries give how much light element $j$ sends to element $i$. The prover commits to $(E,\rho,\text{geometry},B)$ and proves, for every element $i$,
$$\left| B_i - E_i - \rho_i \sum_j F_{ij} B_j \right| \le \epsilon,$$
with $\epsilon$ a public tolerance. It also proves that the image rendered from $B$ under each camera matches the photo within a public photometric bound.
Research Program
The method source is Differentiable Light Transport with Gaussian Surfels via Adapted Radiosity. That paper recasts global illumination over Gaussian surfels (flat disk primitives) as an adapted radiosity solve cached in spherical harmonics, with a center-to-center approximation for the geometric coupling between surfels. It runs at 515 FPS on the Plate of Fruit scene and reports the lowest Chamfer distance on Stanford-ORB. It assumes floating-point GPU execution and has no verification story. We keep its transport model and change what gets computed in the circuit. Their hybrid shooting and gathering solver runs outside the proof, on the prover's GPU. Only the fixed-point residual above and the final rendering are proven.
The circuit has four parts. The first computes form factors in-circuit from committed surfel centers, normals, and areas using the center-to-center formula. Its divisions and inverse square roots use the decompose-then-lookup technique and Goldschmidt refinement of Scalable Zero-knowledge Proofs for Non-linear Functions in Machine Learning, which splits large field elements into 12-bit digits so that lookup tables stay small.
The second part handles visibility, which is the honest weak point. Proving occlusion for every surfel pair would bring back the tree-traversal bloat that makes path tracing unfriendly to circuits. So the prover supplies the visibility mask as a witness. A Fiat-Shamir challenge, meaning randomness derived from a hash of the commitment, selects a random subset of pairs, and for that subset a ray-versus-surfel intersection sub-circuit proves the mask is correct. This gives probabilistic soundness: a prover that lies about a fraction $f$ of visibility entries is caught with probability $1-(1-f)^k$ for $k$ spot checks, which follows from independent uniform sampling. We will measure how large $f$ must be before a cheat actually changes the rendered images, which tells us how large $k$ needs to be.
The third part is the residual and rendering checks. These are sparse matrix-vector products followed by range checks done with lookups. We borrow the Halo2 lookup and layer-reuse machinery of Scaling up Trustless DNN Inference with Zero-Knowledge Proofs, and the random-folding idea that ezDPS uses to compress repeated linear constraints. The fourth part is the relighting check: the buyer's environment map enters as public input, and the circuit evaluates the asset's response under it.
The statistical-audit framing comes from Confidential Guardian. That work proves a hidden classifier's calibration error is below a threshold on a reference set, and it notes the limitation that only the distributions covered by the reference set are audited. Our reference set is the buyer's held-out lighting, so the same coverage caveat applies: a single relighting condition audits one slice of the asset's behavior. Two further papers constrain the design. Stanford-ORB shows that novel-view quality does not certify that lighting was separated from material, which is why our certificate needs a relighting check and not a view check. Radiometric Scene Decomposition shows that specular appearance is very sensitive to small normal errors. For that reason the first version restricts materials to diffuse-plus-low-order spherical harmonics and treats fixed-point drift as a measured risk, not an assumption.
Baselines
The naive reference point is no verification: the buyer receives renders under the training views and trusts them. Against baked assets this baseline should detect nothing. It sets the floor for detection rate and zero for prover cost.
The strongest practical competitor is an unbound relighting test. The buyer sends a secret environment map, the vendor returns a rendered image, and the buyer compares it to a photo under that lighting, as in the Stanford-ORB relighting protocol. Its weakness is that the returned image need not come from the asset being sold. A vendor could produce it with a separate neural relighting model or a better hidden asset. We run it with identical held-out lighting and an identical error threshold, and judge it on two adversaries. The first is a naive baker. The second is a substitution adversary that answers relighting queries from a stronger model while selling a baked asset. Our certificate counts as beating this baseline only if it matches it on the naive baker and catches the substitution adversary, which the unbound test cannot do by construction.
The quality reference is unverified floating-point inverse rendering on Stanford-ORB. For relighting quality, Stanford-ORB reports NVDiffRecMC at a relighting PSNR-H of 24.43, and IDR at a shape Chamfer distance of 0.30. The surfel radiosity paper reports the lowest Chamfer distance but gives no number in our file, so there is no reported result found on Lacuna for its relighting PSNR-H. We will rerun it ourselves in floating point with the authors' settings as the matched reference. Our fixed-point version succeeds if relighting PSNR-H stays within 0.5 dB of that floating-point rerun on the same scenes and splits.
Two ablations isolate the key ingredients. The first replaces the multi-bounce residual with a direct-lighting-only certificate, which tests whether global illumination matters for detection or is just cost. The second replaces the residual check with an unrolled fixed number of Jacobi iterations inside the circuit, which tests whether "verify the solution" really beats "replay the solver" in constraint count and prover time. Both ablations use the same Halo2 backend, the same quantization width, and the same hardware.
Why not just…?
Why not just run the inverse renderer in a trusted execution environment and ship an attestation? For many commercial buyers this is the pragmatic answer, and we treat it as a real alternative, not a strawman. Hardware-Enabled Mechanisms for Verifying Responsible AI Development catalogs the trust it requires: hardware vendor root keys, side-channel exposure, and supply-chain risk. More specific to this setting, an attestation proves that some code ran. It does not give a buyer-checkable statement about the asset's physics unless the attested code itself computes our residual and relighting checks. Our checks are therefore useful either way. The zero-knowledge version matters when there is no shared hardware trust, for example when an asset is resold or audited by a regulator. Our cost measurements will show what that independence costs relative to an enclave.
Why not just verify a held-out relighting image without any cryptography? This is the strongest threat. Against an honest-but-lazy vendor who bakes lighting, it probably works as well as our method, and our experiments are designed to confirm that. The gap is binding. The returned image must come from the committed asset that the buyer receives. The substitution-adversary experiment settles whether this gap is practically important. If buyers can always inspect the delivered asset after purchase, the gap shrinks, and the remaining value of the proof is pre-purchase confidentiality of the asset.
Why not just prove the full differentiable path tracing optimization? Reported costs make this implausible. Native floating-point path-traced reconstruction takes 3–5 hours per object in Shape from Tracing and several minutes per iteration in Inverse Path Tracing. Even a single forward pass of MobileNet v2 takes 2 to 40 minutes to prove in Scaling up Trustless DNN Inference. Proving the optimization also certifies the wrong thing. A buyer cares whether the delivered asset is physically consistent, not whether a particular sequence of gradient steps was taken.
Why not just certify direct lighting only, which is far cheaper? Baked shadows and interreflections are exactly the global-illumination effects that Neural-PBIR and Inverse Path Tracing argue must be modeled to keep albedo clean. Still, a direct-only certificate combined with relighting may catch most baking in practice. The direct-only ablation answers this. If its detection rate matches ours, we drop the multi-bounce residual from the design.
Proposed Experiments
We reconstruct Stanford-ORB objects with the surfel radiosity method and quantize them to 32-, 48-, and 64-bit fixed point. The 64-bit setting follows ezDPS, which reports a 1–2% accuracy drop at that width. We measure relighting PSNR-H and Chamfer distance against the floating-point rerun, which directly tests the precision risk.
For adversaries, we build three families from each honest asset. The first is baked: the training-light rendering is written into the albedo and emission is set to zero. The second is emitter-smuggled: emission is hidden in bright albedo, the ambiguity that Inverse Path Tracing handles with an L1 penalty. The third is the substitution adversary described above. We report detection rate at a fixed false-rejection rate on honest assets, separately for our certificate, the unbound relighting test, and the direct-only ablation.
For cost, we report constraint count, prover time, proof size, and verifier time as functions of surfel count and spot-check count $k$, and compare them to the unrolled-solver ablation. There is no reported result found on Lacuna for proving any rendering computation in zero knowledge. The external reference scale is therefore the zkML literature: 2 to 40 minutes and roughly 6–7 KB proofs for MobileNet v2, and about 333 seconds per data point for ResNet-50 calibration in Confidential Guardian.
The decisive outcome has two parts. Our certificate should catch the substitution adversary, which the unbound test cannot, while losing under half a decibel of relighting quality. And proving a 10^4-surfel object should take time in the same order of magnitude as these zkML figures. If the direct-only ablation detects baking just as well, or if fixed-point drift exceeds the quality bound, the proposal narrows to a direct-lighting certificate. We would report that as the finding.
References
- Differentiable Light Transport with Gaussian Surfels via Adapted Radiosity for Efficient Relighting and Geometry Reconstruction
- Scalable Zero-knowledge Proofs for Non-linear Functions in Machine Learning
- Scaling up Trustless DNN Inference with Zero-Knowledge Proofs
- ezDPS: An Efficient and Zero-Knowledge Machine Learning Inference Pipeline
- Confidential Guardian: Cryptographically Prohibiting the Abuse of Model Abstention
- Stanford-ORB: A Real-World 3D Object Inverse Rendering Benchmark
- Radiometric Scene Decomposition: Scene Reflectance, Illumination, and Geometry from RGB-D Images
- Hardware-Enabled Mechanisms for Verifying Responsible AI Development
- Shape from Tracing: Towards Reconstructing 3D Object Geometry and SVBRDF Material from Images via Differentiable Path Tracing
- Inverse Path Tracing for Joint Material and Lighting Estimation
- Neural-PBIR Reconstruction of Shape, Material, and Illumination