This is the official PyTorch release of MyoFLAME (SIGGRAPH Asia 2026).
Yixuan Shen, Yuyang Ji, Ren Wang, Kaidi Xu, Feng Liu
MyoFLAME is a muscle-driven companion to FLAME. It takes the same parameters as FLAME and tells you which facial muscles produce the expression: 50 named muscle columns (zygomaticus major, frontalis, orbicularis oris, digastric, ...) with an activation in [0, 1] each, and the skin those muscles produce, in FLAME topology. The identity, head, neck, eye and jaw joints stay FLAME's own; only the expression blendshapes are replaced by the muscle response. The jaw muscles drive FLAME's jaw joint: the digastric opens the mouth, the masseter and temporalis close it.
Because the topology, the vertex order and the inputs are FLAME's, MyoFLAME drops into code that already uses FLAME: FLAME 2020 fits from DECA, EMOCA or your own fitting give muscle activations directly, and FLAME landmark embeddings, masks, UV maps and renderers work on MyoFLAME meshes unchanged.
The code was tested with Python 3.12 and PyTorch 2.6 on Linux; a CPU build of PyTorch is enough.
git clone https://github.com/VILab-Drexel/MyoFLAME.git
cd MyoFLAME
python3 -m venv .venv
source .venv/bin/activate
pip install -e .
The MyoFLAME model files (muscle response basis, activation surrogate, jaw calibration) are
part of the repository in myoflame/data/ (61 MB) and need no separate download.
MyoFLAME runs on top of FLAME 2020, which cannot be redistributed. You need to sign up and
agree to the FLAME model license at
https://flame.is.tue.mpg.de, download FLAME 2020 and copy generic_model.pkl into
model/FLAME2020/. MyoFLAME reads the file without chumpy. Other FLAME versions (for example
FLAME 2023, whose expression basis differs) are refused.
import torch
from myoflame import MyoFLAME
model = MyoFLAME("model/FLAME2020/generic_model.pkl") # .to("cuda") for the GPU
# the arguments of smplx.FLAME; all optional, (B, n) each, truncated shape/expression are fine
out = model(betas=shape, expression=expression, jaw_pose=jaw_pose, global_orient=head_pose)
out.vertices # (B, 5023, 3) muscle-driven skin, FLAME topology
out.activations # (B, 50) muscle activations in [0, 1], names in model.muscle_names
out.jaw_open # (B,) jaw opening in radians (0 to 20 degrees)
out.jaw_in_domain # (B,) False where the jaw pose lies outside the calibrated rangeOther FLAME conventions have adapters, so you do not need to reorder pose vectors:
out = model.from_flame_pytorch(shape_params, expression_params, pose_params) # FLAME_PyTorch: pose = [global, jaw]
out = model.from_deca(codedict) # DECA / EMOCA code dictionariesDrive the face with muscles directly, or go back to FLAME parameters:
a = torch.zeros(1, 50)
a[0, model.muscle_names.index("Zygomaticus_Major_L")] = 0.6
a[0, model.muscle_names.index("Digastric_L")] = a[0, model.muscle_names.index("Digastric_R")] = 0.5
out = model.from_activations(a) # smile on the left side, jaw opened 10 degrees by the digastric
flame = model.to_flame(out) # FLAME expression and jaw pose closest to that skinEverything runs in batches on CPU or GPU (about 3 ms per sample on CPU and 0.3 ms per sample
on a GPU in batches of a few hundred) and is differentiable with respect to betas,
expression, the head, neck and eye poses and the activations. Gradients with respect to
jaw_pose flow through FLAME's jaw joint but not through the mapping to physical jaw
coordinates, so they do not reach the activations. activation_mode="exact" replaces the learned surrogate by
the exact inverse-physics solve (CPU, one bounded least-squares problem per sample, not
differentiable); use it for expressions far from the FACS-coded expressions the surrogate
was trained on.
python demos/demo_forward.py --out output/forward
runs FACS-coded expressions of the D3DFACS dataset (shipped as FLAME 2020 parameters) through MyoFLAME and writes, for each, the MyoFLAME mesh, the FLAME mesh and the strongest muscles.
python demos/demo_muscles.py --out output/muscles
drives the face one muscle (or muscle pair) at a time from rest to full contraction and writes OBJ sequences, including the digastric opening the jaw and the masseter closing it.
The muscle columns, the jaw muscle law, the jaw calibration and the file formats are described in docs/MODEL.md. In short: 36 line-muscle columns pull skin towards their bone origin, 8 ring columns (orbicularis oris in four quadrants, orbicularis oculi in its orbital and palpebral parts on each side) contract rings of skin, and 6 jaw-muscle columns open and close the jaw through
q_open = 20 deg * clip(mean(digastric) - 5.23 mean(masseter) - 2.10 mean(temporalis), 0, 1)
where 5.23 and 2.10 are the closers' torque capacities relative to the jaw-depressor group. The activations are estimates of how much each muscle contributes to the observed FLAME deformation, not measurements of muscle effort.
python -m unittest discover -s tests -v
Tests that need FLAME skip until model/FLAME2020/generic_model.pkl is present.
If you use MyoFLAME, please cite
@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},
}
