Publications

GNM: an open parametric head model covering face, neck, eyes, teeth, and tongue, built from high-res scans and artist data.

SHELLS reconstructs topologically consistent 3D heads from multi-view images in 0.08 s via layered surface sampling.

CUBE: a 3D face representation that combines B-spline volumes with learned features, for scan registration and monocular reconstruction.

One diffusion model that generates 3D face geometry from sketches, photos, landmarks, FLAME parameters, or text.

A diffusion portrait generator that jointly learns RGB and depth, enabling depth estimation and depth-driven editing.

Recovers studio-quality face geometry and reflectance maps from a single in-the-wild video of a head rotation.

ScaffoldAvatar couples local patch expressions with 3D Gaussian splatting for ultra-high-fidelity photoreal head avatars.

NFDT transfers blendshapes and animations from a template character to a new character given only its neutral shape.

Transformers without positional encoding: sequences become smooth latent spline trajectories with editable control points.

A generalized physical face model learned from data that fits any new identity into a ready-to-animate simulation rig.

Architectural changes that make continuous 2D facial landmark detectors more accurate and temporally stable.

Neural head avatars that can be both animated with unseen expressions and relit in any environment.

Implicit networks that jointly model skin and underlying anatomy, a drop-in replacement for blendshape rigs.

Animates skin micro-wrinkles that widen and deepen under stretch and compression, with a real-time shader approximation.

Style transfer from real skin scans onto simulated wrinkles, giving controllable yet organic micro-detail across whole faces.

An implicit neural physics face model that can be driven independently by expression and by per-subject style.

A perceptual loss for monocular 3D face reconstruction that scores shape by how well its shading matches the image.

ReNeRF: relightable radiance fields captured with a few area lights, supporting novel views and near-field lighting.

Graph-based simulation that synthesizes realistic skin micro-wrinkle geometry, scaling to millions of wrinkles.

The first facial landmark detector that predicts any number of continuous landmarks, queried by 3D position at inference.

FRAN: production-ready, temporally stable face re-aging for video, trained on synthetic longitudinal data.

Video face swapping by jointly learning dynamic 3D geometry and texture with identity-specific decoders.

A transformer autoencoder for 3D facial animation that disentangles identity from motion for synthesis and retargeting.

MoRF turns NeRF into a generative head model, rendering multiview-consistent heads with controllable, morphable identity.

The first multi-view pipeline that tracks dense 3D facial hair and the skin beneath it throughout a performance.

Retargets captured facial performances to digital characters using local expression transfer plus anatomical constraints.

Shows how more physically accurate lighting models (near-field lights, indirect bounces) improve facial appearance capture.

A transformer-based 3D shape model that is trained once and works on any mesh topology, for faces, hands, and bodies.

Completes high-quality skin-only renders with neural rendering to produce photoreal full-head portraits without artist work.

AdaConv generalizes AdaIN to transfer both statistical and structural style, enabling real-time structure-aware style transfer.

A nonlinear neural 3D face model with disentangled, semantic control over identity and expression.

Attention-driven cropping lets landmark detectors run on very high-resolution face images without exhausting GPU memory.

Tracks road edges from a vehicle-mounted camera using particle filters, for advanced driver assistance systems.

Segments exudates in retinal fundus images and grades diabetic retinopathy severity using error-boost feature selection.