Physics-based face models work from the inside out. They place muscles, bones and fat under the skin and simulate how the tissue moves. The most general recent versions fit a new person by warping one shared "canonical" anatomy, learned from a modest capture cohort, until it matches that person's scan. They then drive expressions with actuation patterns shared across identities. This design builds a default anatomy into the model. The bias-amplification literature shows that defaults learned from skewed data tend to become the model's most likely output rather than a small tail effect. This proposal asks whether the same thing happens in faces. When the model animates someone whose demographic group was underrepresented in its training cohort, does the generated smile, frown or pucker move like that person, or does it move like the majority group's version of it?
The question matters because these models are being sold on zero-shot generality: animating "hundreds of identities from single scans." If that generality quietly replaces a person's own motion with majority-group motion, the failure is invisible in pooled error metrics. It would surface in exactly the users the model was meant to generalize to. Small departures from natural motion are not harmless, either. In a perception study of animated digital humans, intensity outside natural limits measurably lowered perceived sincerity and trustworthiness (Microexpressions in digital humans).
Core Claim
We claim that the pull toward the majority group is a measurable, controllable quantity. In anatomically grounded animation it has two separable sources. The first is the expression model: how actuation is shared across people. The second is the identity fit: how hard the canonical anatomy resists being warped. We will measure it in a controlled setting where we choose the demographic skew of the training cohort. We will compare the physics model against a surface model given identical data. We will then show which ingredient of the physics model is responsible.
The new part is not "physics models have subgroup error gaps." Pooled per-group error is easy to report and says nothing about direction. The new part is a directional measurement. The proposal adapts the directional bias-amplification idea of Directional Bias Amplification from discrete classification to continuous 3D motion. It asks whether generated motion is more majority-like than real motion, and by how much. The main ingredient being tested is the single canonical material space.
Positioning
Two physics papers supply the method and the gap. Learning a Generalized Physical Face Model From Data trains a network that warps one canonical cranium, jaw and soft-tissue space to fit 336 identities using physically inspired losses. An Implicit Physical Face Model Driven by Expression and Style maps identities into a shared canonical space and separates expression from personal style, using nine identities. Neither paper reports anything broken down by demographic group, and both are built on small or single-source cohorts.
High-Quality Face Capture Using Anatomical Muscles supplies the readout that makes the physics side worth studying. Its 60-muscle differentiable simulation reproduces expressions with roughly 12 active muscles, where a blendshape rig needs more than 100 interacting weights. This sparsity lets us ask which anatomically named muscles carry the majority pull. A surface rig cannot answer that cleanly.
On the bias side, Directional Bias Amplification supplies the measurement logic. It also supplies a constraint, its masking paradox: removing the attribute from the input can increase inference of that attribute from context. FairBelief motivates the hypothesis by showing that stereotyped associations sit at the peak of the model's likelihood, which is the role a canonical template plays geometrically.
Investigating Bias and Fairness in Facial Expression Recognition is the closest neighbour. It studies 2D expression recognition and finds that data augmentation raised accuracy (by about 7%) without fixing subgroup bias, while adversarial disentanglement helped. We differ in three ways: we study generation rather than recognition, 3D rather than 2D, and the direction of error rather than error size.
We deliberately set aside the tempting "LLM says men are angry, so the avatar frowns" cascade built on Angry Men, Sad Women. In that pipeline the bias lives in the emotion label. A faithful animator would simply render it, and the muscle layer adds no evidence.
Research Program
The testbed is FaceScape: 938 subjects, each scanned in 20 fixed expressions, with topologically uniform meshes so displacements are in correspondence. Its subjects are primarily from Asia. That rules out race as a controlled axis, but it is convenient for a controlled study, because we can manipulate other axes (sex and age band) without confounding them with capture setup. We will use released subject metadata where it exists. Otherwise we will collect perceived sex and age-band annotations and report annotator agreement.
The task is expression prediction from a neutral scan. Given only a held-out subject's neutral face, the model predicts that subject's 19 non-neutral expressions. We score the predictions against the real scans. We build training cohorts at controlled majority shares $p \in \{0.5, 0.7, 0.9\}$ along one axis at a time, keep cohort size fixed, and always evaluate on a balanced held-out set.
We use two directional measurements. Let $d_{s,e}$ be the true displacement of subject $s$ in expression $e$ relative to their neutral scan. Let $\hat d_{s,e}$ be the prediction, and $\Delta_e = \mu^{\text{maj}}_e - \mu^{\text{min}}_e$ the difference between real group-mean displacements, estimated on held-out data. The majority pull for a minority subject is
$$ \beta = \frac{\langle \hat d_{s,e} - d_{s,e},\ \Delta_e \rangle}{\|\Delta_e\|^2}, $$
the fraction of the gap between groups that the model's error closes in the majority's direction. A value of $\beta = 0$ means no directional pull. A predictor that outputs the pooled training mean gives $\beta \approx p$, which sets the "pure default" reference.
The second measurement transfers the attribute-inference half of the directional framing. We train a classifier only on real displacement fields, never on face shape, to predict group from how the face moves. We then report the increase in the "majority" classification rate on generated minority motion relative to real minority motion. A positive gap means the model has written the training skew into its output dynamics. This is amplification in Wang and Russakovsky's sense: the association between group and motion in the outputs is stronger than in the data.
The physics model gives us a third, mechanistic readout. We fit muscle activations to each prediction and compare them per group with activations fitted to the real scans. This shows whether the pull is concentrated in particular muscles, for instance around the mouth. We also report per-group physical plausibility: lip volume change in pucker and lip–teeth interpenetration. Together these test whether the canonical anatomy's warp is more strained for minority faces.
To isolate the ingredient, we vary only how the identity is represented. One version uses the single canonical material space. Another uses a canonical space mixed from several group-specific templates. A third weakens the regularizer that keeps identity codes close to the canonical, swept over strength. If $\beta$ tracks the canonical prior rather than the expression decoder, the canonical space is the culprit.
Baselines
The naive reference is a pooled-mean transfer. It adds the training cohort's mean displacement for each expression to the test subject's neutral face. It ignores identity entirely, so it fixes the "pure default" end of the scale ($\beta \approx p$) and the ceiling of the classifier gap. Any learned model that does not clearly beat it on $\beta$ is not personalizing at all.
The surface baseline is FaceScape's own riggable model, the bilinear identity–expression model from FaceScape. We retrain it on each skewed cohort. It is the strongest competitor on this question because it is the representation a skeptic would use instead: it is purely statistical and has no anatomy prior. It therefore tells us whether physics adds pull, removes it, or does nothing.
The strongest existing method is the generalized physical face model of Learning a Generalized Physical Face Model From Data. We reimplement it from its stated losses (skin reconstruction, skull fixation, bone rigidity, biomechanical plausibility), using SCULPTOR-style bone constraints as in that paper. That paper gives no numeric reconstruction error, and no reported result was found on Lacuna for any method on FaceScape neutral-to-expression prediction. The closest physics number, 0.3849 mm vertex error for An Implicit Physical Face Model Driven by Expression and Style, comes from a different nine-identity dataset and is not comparable. We will therefore establish all reference numbers ourselves.
All models get the same training subjects, the same neutral-scan inputs at test time, the same correspondence topology, and an equal hyperparameter search budget. Each search is tuned on a balanced validation split, so no model is tuned to the majority.
The key ablations are the three identity representations above: single canonical, mixed canonical, and swept regularization. Each is crossed with a reweighted-data variant. "Beating" the baselines means reducing $\beta$ and the classifier gap for minority subjects at $p = 0.9$ by a margin outside the subject-bootstrap confidence interval, without raising majority-group vertex error by more than that interval. A result where physics shows less pull than the surface model, with the gap explained by the canonical ablation, is equally informative and would be reported as such.
Why not just…?
Why not just use blendshapes and skip the muscles? We do include a surface model, and if it shows the same pull with the same ablation behaviour, that is a legitimate finding that physics is irrelevant to this bias. The argument for physics is not visual quality. It rests on two concrete points: the canonical anatomy is a distinct, nameable prior that surface models lack, and the roughly 12-muscle sparsity reported in High-Quality Face Capture Using Anatomical Muscles lets us locate the pull anatomically rather than across 100 interacting weights.
Why not just rebalance the training cohort? This is the most serious threat. We test it directly with reweighting and with the $p = 0.5$ cohort. If balanced data drives $\beta$ to the balanced-cohort floor, the conclusion is that collecting diverse captures suffices, and that is worth knowing. We expect rebalancing to be insufficient for one reason. The single canonical space is a single anatomy regardless of sampling weights, and in 2D recognition augmentation alone did not remove subgroup bias (Investigating Bias and Fairness in Facial Expression Recognition). Whether that carries over to 3D generation is precisely what the experiment settles.
Why not just subtract a per-group offset after generation? This is the continuous analogue of the threshold recalibration that cut measured amplification from 0.040 to 0.001 in Directional Bias Amplification. We implement it as a baseline. It needs group labels at inference time, and it corrects only the group mean. Anatomizing Bias in Facial Analysis calls global post-hoc fixes a trap that hides failures in the learned representation. If our classifier gap and per-subject $\beta$ variance survive mean-correction, the offset fix is shown inadequate; if they do not, it wins.
Why not just hide demographic cues from the input? The masking paradox (Directional Bias Amplification) predicts that the model will infer group from face shape anyway. We test this by training the expression model on shape-normalized neutrals. We then check whether $\beta$ falls, or whether the classifier gap rises.
Proposed Experiments
The first decisive check is the dose–response curve: $\beta$ and the classifier gap plotted against $p$, for each model, along the sex axis and the age-band axis separately. Inheritance shows up as a positive slope. Amplification shows up as a positive classifier gap, meaning generated motion reads as more majority than real motion does.
The second check is attribution. If the single-canonical physics model's pull drops under the mixed-canonical or weak-prior variants while vertex error holds, the canonical anatomy is causally implicated. If nothing changes, the pull lives in the shared expression actuation instead.
The third check is mechanistic. We compare fitted muscle activations and the lip-volume and interpenetration metrics per group, to see whether minority faces show both more strained tissue fits and more pull, as the canonical-warp hypothesis predicts.
As a small confirmation, we will run a perception study on matched renders of minority subjects. Observers will rate the subject's own real expression against the predicted expression from the high- and low-pull models. The aim is to test whether larger $\beta$ predicts lower perceived naturalness and sincerity, in line with Microexpressions in digital humans.
Race and intersectional axes cannot be studied in FaceScape. We state this as the main limitation. The measurement protocol is designed to transfer unchanged once a demographically broader capture set exists.
References
- Microexpressions in digital humans: perceived affect, sincerity, and trustworthiness
- Directional Bias Amplification
- Learning a Generalized Physical Face Model From Data
- An Implicit Physical Face Model Driven by Expression and Style
- High-Quality Face Capture Using Anatomical Muscles
- FairBelief - Assessing Harmful Beliefs in Language Models
- Investigating Bias and Fairness in Facial Expression Recognition
- Angry Men, Sad Women: Large Language Models Reflect Gendered Stereotypes in Emotion Attribution
- FaceScape: A Large-Scale High Quality 3D Face Dataset and Detailed Riggable 3D Face Prediction
- Anatomizing Bias in Facial Analysis