Neural face riggers now take an arbitrary head mesh, with any number of vertices, any triangulation, and separate eyeball and teeth pieces, and output a full set of animation controls in seconds. Their key promise is that the result does not depend on how the surface happens to be cut into triangles. Nobody actually checks that promise on the meshes where it matters. When a studio feeds a new sculpt from Objaverse or CGTrader into such a system, there is no ground-truth rig to compare against, so there is no way to tell whether the output is trustworthy. We propose a cheap check that needs no labels. Re-cut the same surface several ways: remesh it, change its resolution, split it along fake seams, or weld its separate pieces together. Run the rigger on each version, and compare the predicted motion at a few randomly sampled surface points. If the rigger were truly independent of triangulation, the answers would agree. Where they disagree, at least one of them is wrong.
The idea borrows its shape from interactive proof theory, where a cheap verifier asks an untrusted, powerful party a handful of random questions. Local agreement on those answers then certifies a global property, and the verifier's cost scales with the confidence it wants, not with the size of the object. We do not import any quantum machinery. We import a style of argument. The core claim is that sampled cross-discretization disagreement is a label-free, size-independent estimate of how much a neural rig depends on triangulation. We further claim that this estimate predicts true rig error per mesh well enough to flag failures before an artist sees them. Used as a training signal, it should also reduce the failures that existing systems document on unusual meshes.
Where the gap is
Three systems define the current state of topology-agnostic facial deformation. All three build on DiffusionNet, a surface network that passes information by simulated heat flow, which in principle makes it insensitive to how the surface is triangulated.
Neural Face Rigging predicts a per-triangle deformation gradient (a "Neural Jacobian Field") driven by 53 interpretable expression controls. It requires a manual step that strips out the mouth interior and eyes. RigAnyFace adds a global encoder so that disconnected pieces move together, and it trains on unlabeled sculpts using 2D motion supervision. It is reported to be sensitive to extreme discretizations and to fail on "shell-like" heads that lack an inner surface. ScanTalk drives unregistered scans from speech, and it stretches lips when the mouth surface is closed.
Each of these papers motivates our proposal: each one reports failure modes tied to how the mesh is built rather than to what shape it represents. None of them measures discretization dependence directly. Their benchmarks evaluate on clean, registered meshes, which is exactly where the dependence cannot show up.
The verification literature supplies the method. In Quantum soundness of testing tensor codes, a global object passes a random local agreement test only if it is close to a genuinely consistent object. The paper also adds a second, structurally different test (a commutation check), because plain agreement tests miss certain cheating strategies. Classical Verification of Quantum Proofs shows a verifier using logarithmic-size questions and constant-size answers. Classical Commitments to Quantum States makes verifier work scale with a security parameter rather than with the size of the state.
The Power of Unentanglement shows that, under a conjecture, any number $k$ of independent provers can be simulated with two. This gives us a design constraint rather than a theorem to use. We ask whether comparing two discretizations is enough, or whether $k$ are needed. The answer turns out to be elementary in our setting.
The research file found no existing work connecting interactive verification to geometric deformation. The closest graphics practice is soft regularization, such as as-rigid-as-possible penalties. That literature reports such penalties are easily overpowered by strong driving losses (Kurenkov et al., Kim et al.), which constrains our design: the test must work as a measurement, not only as a penalty.
Core Claim
Let $S$ be a head surface and let $M_1, \dots, M_k$ be discretizations of it produced by known operations, so that every surface point $p \in S$ has an exact location on each $M_i$ via barycentric coordinates. Let $f_i(p, z)$ be the deformed position of $p$ predicted by the rigger on $M_i$ under expression code $z$.
The disagreement score is
$$D = \mathbb{E}_{i \ne j}\, \mathbb{E}_{p \sim S}\, \mathbb{E}_{z \sim Z}\, \lVert f_i(p,z) - f_j(p,z) \rVert ,$$
where $p$ is sampled uniformly by area and $z$ is drawn from a fixed distribution over expression controls. Two simple facts give the test its guarantees.
The first is query cost. Each term in the average is bounded by the face's scale, so Hoeffding's inequality says that $q = O(\varepsilon^{-2}\log(1/\delta))$ sampled triples estimate $D$ to within $\varepsilon$ with probability $1-\delta$. That cost is independent of vertex count.
The second is that two copies suffice. Let $\bar f$ be the average prediction across discretizations. The triangle inequality and convexity of the norm give $\mathbb{E}_i \lVert f_i - \bar f \rVert \le D$. Small pairwise disagreement therefore certifies that a typical discretization is close to a single consensus field, which is the "two provers suffice" lesson in miniature.
What $D$ cannot certify is that the consensus is correct. A rigger that is consistently wrong passes the test. The empirical question, and the heart of the proposal, is how often real failures on in-the-wild meshes are inconsistent failures.
Research Program
We will build a perturbation suite with five families of discretization change, all of which preserve the underlying shape.
The first two families are the generic ones. We remesh the surface isotropically at three resolutions, and we apply anisotropic, sliver-heavy retriangulation, which targets RigAnyFace's sensitivity to extreme discretizations.
The other three families target topology:
- Seam splitting. We cut the surface along random curves that are invisible geometrically, producing duplicated vertices and disconnected pieces. This probes whether the global encoder's coordination actually depends on connectivity.
- Component welding. We merge eyeballs or teeth into the face shell at their nearest contact, which is the inverse operation.
- Shell thickening or thinning. We add or remove an inner backing surface, which targets the shell-like failure.
The welding and splitting families play the role of the commutation test in the tensor-code work. They are a second, structurally different check that catches failures the plain remeshing agreement test misses. We will measure whether that is so.
On top of the suite we build two things.
The first is a per-mesh reliability score, $D$ estimated from a few hundred queries. It is computed separately per perturbation family and per facial region (lips, lids, brows, interior components), so that a flag says where and why.
The second is a training variant. During training, a pair of perturbed discretizations is drawn and the sampled disagreement is penalized. This is compared head-to-head against plain remeshing augmentation (described under Baselines).
Every model is evaluated with its released weights where available. Where weights are not available, we reimplement it on the same data, and we report which case applies.
Baselines
The naive reference point is to deploy the rigger with no reliability signal. For flagging, this corresponds to a random ranking of meshes. For training, it corresponds to each model's standard recipe. The strongest existing rigger is RigAnyFace, which reports 1.01 mm blendshape reconstruction mean absolute error on humanoids against production rigs, versus 2.77 mm for Neural Face Rigging, on its curated benchmark. For speech-driven scans, the reference is ScanTalk, with a lip vertex error of 3.012 on VOCAset.
The strongest competing reliability signal is deep-ensemble disagreement: five RigAnyFace models trained from different seeds, evaluated on the original mesh, with the spread of their predictions as the score. It is the standard label-free uncertainty estimate and needs no geometric insight, so it is the competitor our score must beat. It gets the same training data. Our score gets one model run on five perturbed meshes, so the two signals use matched inference compute. We beat it if the Spearman correlation between our score and true per-mesh mean absolute error is higher on held-out in-the-wild heads, with a bootstrap confidence interval excluding zero difference. We also beat it if, at a fixed flag rate of 10%, our score captures more of the worst-error meshes.
The cheapest geometric competitor is a validity check: counting inverted triangles and running bounding-volume-hierarchy self-intersection tests on deformed frames. It targets a different failure, so we test whether $D$ adds predictive power beyond it.
For training, the competitor is remeshing augmentation with a dense consistency loss on corresponding vertices, given the same perturbation suite and tuning budget.
Ablations isolate the key ingredient in two ways. We drop each perturbation family in turn, in particular the topology families versus remeshing alone. We also vary the number of discretizations compared, from two up to five, to test whether two suffice in practice as the bound suggests.
Why not just…?
Why not just use ensemble disagreement? Ensembles measure sensitivity to training randomness. Every ensemble member sees the same triangulation, so a shared bias toward connectivity artifacts, such as lips welded shut by a closed mouth surface, makes them agree confidently. That is our hypothesis, not established fact. The matched-compute comparison above settles it, and if ensembles win, the proposal reduces to a negative result about geometry-specific testing that is worth reporting.
Why not just augment training with random remeshing? This is a real threat to the training half of the proposal. Our consistency loss on point pairs is close to it, and we expect much of any training gain to come from augmentation alone. We therefore claim novelty mainly for the deployment-time test and for the topology-changing perturbation families, which ordinary remeshing augmentation does not include. If remeshing augmentation matches our training variant, we report that and keep the test.
Why not just run exact collision and inversion checks? These checks are already fast with standard spatial data structures. The quadratic cost quoted in the research file is a worst case, so we do not argue speed. These checks detect invalid geometry, but most rig errors, such as a brow that does not rise or teeth that drift, are valid geometry in the wrong place. The incremental-correlation experiment measures how much $D$ adds beyond them.
Proposed Experiments
Experiment 1: measurement. We report $D$ for Neural Face Rigging, RigAnyFace and ScanTalk under each perturbation family on Multiface, on ICT FaceKit identities, and on the RigAnyFace in-the-wild set. Region-wise scores tell us whether the documented failures (closed mouths, shell heads, disconnected eyes) show up as discretization inconsistency.
Experiment 2: the decisive check. On heads with production rigs, we correlate per-mesh $D$ with true reconstruction error and run the flag-rate comparison against ensembles and validity checks. If the correlation is weak, then failures are mostly consistent and the central claim fails.
Experiment 3: training. We retrain RigAnyFace with the consistency loss. Success means held-out mean absolute error at or below 1.01 mm on humanoids, together with lower $D$ under topology perturbations, compared with the remeshing-augmentation baseline. On VOCAset and BIWI, we check that lip vertex error stays at or below 3.012 when ScanTalk is retrained the same way. The research file reports no BIWI number to compare against.
Experiment 4: copies. We sweep the number of discretizations from two to five and confirm that ranking quality saturates at two or three. This is the empirical counterpart of the two-copies bound.
References
- Neural Face Rigging for Animating and Retargeting Facial Meshes in the Wild
- RigAnyFace: Scaling Neural Facial Mesh Auto-Rigging with Unlabeled Data
- ScanTalk: 3D Talking Heads from Unregistered Scans
- Quantum soundness of testing tensor codes
- Classical verification of quantum proofs
- Classical Commitments to Quantum States
- The Power of Unentanglement
- DeformNet: Free-Form Deformation Network for 3D Shape Reconstruction from a Single Image
- As-Plausible-As-Possible: Plausibility-Aware Mesh Deformation Using 2D Diffusion Priors