MyoFLAME

Inverse-Physics Muscle Activation for Statistical 3D Face Models
Yixuan Shen, Yuyang Ji, Ren Wang, Kaidi Xu, Feng Liu
SIGGRAPH Asia 2026 Conference Papers

Rendered live in your browser at the FLAME mean identity: for six FACS-coded D3DFACS expressions the surrogate gφ predicts the 50 activation coordinates from the FLAME expression and jaw pose, and the skin is the muscle response of those coordinates (the three largest activations are listed; jaw opening comes from the calibrated physical jaw map). The last head is a single hand-set muscle, which the basis handles one side at a time.

Abstract

Parametric 3D face models such as FLAME have become a common interface for monocular face regression, performance capture, facial animation, and avatar systems. While compact and widely adopted, these models describe how the facial surface deforms but not which anatomical structures explain that deformation: a captured smile carries no information about whether it is best explained by Zygomaticus Major or Risorius, or whether its activation pattern is bilaterally symmetric in muscle space. We introduce MyoFLAME, a drop-in biomechanical companion to FLAME that adds an anatomical interpretation layer to any FLAME-fitted face while preserving the FLAME interface end-to-end. Given any FLAME (β, ψ, pose) input, MyoFLAME produces a coordinated three-layer anatomical output: 42 named muscle activation coordinates, the supporting skull mesh, and the matching skin surface, aligned with FLAME's own output. These coordinates are defined by a constrained inverse-physics projection of FLAME-derived surface displacement onto a fixed muscle response basis rather than by in vivo measurements of physiological muscle effort. Our approach combines a hybrid physical response basis (linear-elastic finite-element responses for bone-to-skin muscles with analytic ring-contraction responses for sphincters, precomputed once on a canonical template) with closed-form linear maps from FLAME identity coefficients to per-subject skull and muscle-attachment geometry. Per-sample inverse-physics optimization is amortized into a compact residual surrogate, delivering activation-coordinate inference at interactive rates.

This page runs MyoFLAME 1.0 in your browser: 50 named muscle columns (line muscles, the orbicularis oris in four quadrants, the orbicularis oculi in its orbital and palpebral parts) and six jaw muscles that open and close the jaw through the jaw muscle law. Everything below is computed live from the files in viewer/data; nothing is pre-rendered.

Interactive model viewer

Like the FLAME viewer, the left panel drives the model and the right canvas shows the FLAME-topology head (5,023 vertices). Two ways to drive it: FLAME expression mode takes identity β, expression ψ and jaw pose exactly as FLAME would, runs the surrogate gφ (the same weights as the released checkpoint, ported to JavaScript) to obtain the 50 activation coordinates, and renders the muscle response in place of the expression blendshapes; Muscles mode lets you set the 50 activations yourself. The jaw sliders move along the calibrated physical jaw manifold (opening and lateral excursion), so the muscle response grid is interpolated the way the runtime does it. The overlay options show FLAME's own blendshape expression for the same input and the per-vertex distance between the two.

Identity shape (±2σ, LYHM) hide

FLAME expression (±2σ, D3DFACS) hide

Jaw and neck hide

Muscle activations (50) hide

Display hide

Lower the skin opacity to see the layers below.
1.00

Predicted activations hide

loading model data ...
mm
08

What is exact and what is approximate here. The skin is computed with the same arithmetic as the released runtime (FLAME blend shapes and linear blend skinning, bilinear interpolation of the 5x5 jaw response grid, LBS transport of the response vectors). The response columns shown by default are a display repair of the released basis, made once at export and documented in the data manifest: the released FEM solve leaves node-level spikes in a few columns (up to 70 mm per unit activation on single brow and nose vertices) and gives no response to the vertices inside the mouth, which tears the lips and cheeks; the repair replaces spike vertices by their ring average, extends the field harmonically to the unsupported mouth vertices and applies two passes of Laplacian smoothing to the displacement field (mean change over the observed vertices mm per unit activation at the closed jaw; the per-muscle change is in the table below). The raw release basis switch shows the unrepaired columns at the closed jaw for comparison. On load the page re-runs reference cases computed by the Python package and prints the largest deviation in the status line. The response grid and blend shapes are stored in 16-bit floats (largest rounding error of a response column: mm). The jaw manifold is a 65x65 table of the bicubic calibration map. The muscle layer under the skin is the model's own definition drawn literally: each column's footprint as a sheet 1.2 mm under the skin (it deforms with the skin and brightens with activation), each line muscle as a fusiform belly from its bony origin (skinned with the skull, so mandible origins follow the jaw) to its displaced skin insertion (it shortens as the insertion moves and bulges with activation), and each sphincter as a ring threaded through its own region vertices. The skull is the template skull and is not adapted to the identity sliders.

The 50 muscle columns

Thirty-four Type-A muscles are bone-to-skin line muscles: each pulls a footprint of the FLAME skin along the line from its skin insertion to its bony origin (mesh-geodesic Gaussian across the line, exponential decay towards the origin), with the full three-dimensional pull where the insertion sits on soft tissue and the tangential pull only where it sits on bone (brow, nose, upper eyelid). Eight Type-B columns are ring contractions of FLAME regions: the orbicularis oris in four quadrants (upper and lower lip, left and right; the quadrant weight is a smooth tanh across the midline so the left and right quadrants join smoothly) with a forward (protrusion) share, and the orbicularis oculi split into an orbital part (AU6 cheek raiser: the ring outside the lids) and a palpebral part (AU7 lid tightener: the lid band within 3 to 7 mm of the globe, pulling tangentially to the eyeball). The levator palpebrae superioris (AU5) raises the upper lid only; the platysma (AU20/AU21) pulls the lower face down and back along the neck. Six Type-C jaw muscles (digastric, masseter, temporalis, left and right) exist in the basis but their coordinates are fixed to zero at inference because jaw rotation is carried by FLAME's own jaw joint. All columns pass through one screened-Poisson surface operator (16 mm smoothing length, lip-contact springs, unilateral lip contact) and carry a mirror-symmetric correction drawn from FLAME's expression space inside their footprint, limited so that no single column and no reconstruction of the 218 D3DFACS expressions tears or flips a face. Strengths are hand-defined per muscle (a = 1 moves the most displaced vertex by the listed millimetres). The table lists each column's neutral-jaw displacement scale (RMS over the 3,931 observed skin vertices for activation 1.0), its footprint width and strength, and the anatomical note of its landmark table.

Click a muscle name to set it to 0.5 in the viewer (Muscles mode) and highlight its fibres.

Citation

@inproceedings{shen2026myoflame,
  title = {MyoFLAME: Inverse-Physics Muscle Activation for Statistical 3D Face Models},
  author = {Shen, Yixuan and Ji, Yuyang and Wang, Ren and Xu, Kaidi and Liu, Feng},
  booktitle = {ACM SIGGRAPH Asia Conference on Computer Graphics and Interactive Techniques (SIGGRAPH Asia)},
  year = {2026},
  doi = {10.1145/3829340.3842189},
}

Code and MyoFLAME model files are released under the MIT licence. FLAME (Li et al., SIGGRAPH Asia 2017) keeps its own licence and is not part of this repository: demos/build_viewer_data.py writes the viewer's FLAME data from your own FLAME 2020 file. The D3DFACS expressions are FLAME parameters averaged over the D3DFACS registrations of each FACS code.