With the growing demand for realistic virtual humans, parametric body models have become a cornerstone of modern medicine, sports, and entertainment applications. However, most of these models are inherently limited: they only capture the 3D surface of the skin, offering no insight into the complex bio-mechanical structures that generate motion. As more applications expand towards biomechanics, the need for virtual human models that go beyond the skin has become increasingly evident. Traditional soft-tissue simulations, such as FEM, are accurate but non-scalable and too computationally expensive for most common applications. Alternatively, existing biomechanical tools can simulate muscular forces and activations, but do not model changes in external shape, restricting how activations correlate with actual observable anatomy. This motivates a novel inverse research problem: recovering muscle deformations directly from visible surface observations — i.e., from the skin, and thus the pose. In this work, we present SOMA (from Surface Observations to Muscle Anatomy), a person-specific model that infers spatio-temporal muscle behavior from surface signals obtained using RGB cameras, and SKIM, a subject-specific soft-tissue deformation dataset. To the best of our knowledge, this is the first method that attempts to recover muscle deformations from multi-view RGB data. We show how our method provides anatomically grounded animations without the complexity of traditional simulations, leading to a scalable and cost-effective solution. Data and code are available.
SOMA builds on the idea of pose-dependent corrective blendshapes, but extends them from a single skin surface to a volumetric, multi-layer anatomy. Given a skeletal pose, two cascaded non-linear U-Nets predict a muscle displacement field Dmusc that drives the bulging of the underlying muscle layer, and a residual offset Dres that lets the skin slide and compress over the tissue instead of rigidly following the muscle. Because recovering these layers from sparse surface markers alone is ambiguous, we supervise the network with the canonical marker residuals from SKIM and regularize it with a set of biomechanically inspired priors: area-normalized Laplacian smoothness and biharmonic bending resistance, edge-stretching and tangential-sliding constraints, and a prism-based volume-preservation term that enforces the near-incompressibility of soft tissue. Finally, the learned boundary deformation is propagated to high-resolution individual muscle meshes through a precomputed barycentric binding, yielding anatomically grounded, spatio-temporal muscle animation directly from pose.
Skin Layer
Muscle Layer
Skeleton Layer
Individual Bones
Inferring internal muscle deformation from RGB video is inherently ill-posed, since only the skin surface is observed. SKIM resolves this by pairing high-precision, temporally consistent skin annotations with an individualized volumetric muscle template. Each of the five subjects wears a custom skin-tight suit embedded with ArUco markers and is captured in a 120-camera markerless motion-capture studio, while a separate 140-camera scanner reconstructs high-resolution static templates of the skin, muscle, and skeleton layers, together with individual muscle meshes. Detected markers are unwrapped into a canonical point cloud, bound to the underlying muscles, tracked across motion, and converted into pose-normalized residual deformation fields that capture subject-specific soft-tissue dynamics. In total, SKIM provides 45 minutes of multi-view recordings with ground-truth skeletal poses, marker trajectories, visibility masks, and the full multi-layer anatomy for each subject.
From a single skeletal pose, SOMA recovers how each muscle deforms beneath the skin. For three held-out motions, we show the predicted skin layer, the aggregated muscle layer, the individual muscle meshes, and the per-vertex deformation heatmap, all synchronized in time.
Our multi-layer formulation produces a coherent anatomy stack — skin, muscle, and bone — that deforms consistently. Below we show every layer of a single sequence side by side, in both textured and deformation-heatmap renderings.
Skin Layer
Skin Heatmap
Muscle Layer
Muscle Heatmap
Individual Muscles
Individual Muscles Heatmap
Skeleton Layer
Individual Bones
@inproceedings{alvaradosoma2026,
author = {Alvarado, Eduardo and Kim, Emily and Nolte, Gerrit and Runte, Friedemann and Botsch, Mario and Habermann, Marc and Theobalt, Christian},
title = {SOMA: From Surface Observations to Muscle Anatomy},
booktitle = {European Conference on Computer Vision (ECCV)},
year = {2026},
}