SOMA

SOMA implements SOMA-X identity, pose, and corrective controls without requiring py-soma-x.

Setup

SOMA downloads automatically on first use from the abcamiletto/body-models Hugging Face repository, which records the SOMA-X Apache 2.0 provenance. To prefetch:

body-models download soma

The loader requires normalized assets, as provided by the hosted archive. To normalize upstream SOMA-X 0.2.1 assets and save their path:

body-models preprocess-soma /path/to/upstream /path/to/processed

SOMA exposes 77 public joints and uses an internal twist-joint rig for skinning. The lod options "mid", "low", and "xlo" have 18,056, 4,505, and 612 vertices, respectively.

prepare_identity() defaults to repose=True, bind_pose="fit", matching SOMA-X. Use repose=False to retain the fitted rest shape and skeleton, bind_pose="fit_detached" to stop gradients through fitting, or bind_pose="canonical" for the canonical bind pose.

API

body_models.soma.numpy.SOMA

SOMA(
    *,
    model_path=None,
    model_type="soma",
    lod="mid",
    rotation_type="axis_angle",
    simplify=1.0,
)

Bases: body_models.soma._model.SOMA

Native SOMA-X model with identity, pose, and corrective controls.

METHOD	DESCRIPTION
apply_pose_correctives	Apply prepared pose correctives to identity-dependent rest vertices.
forward_points	Compute positions defined by a prepared vertex mapping.
forward_skeleton	Compute posed public-joint transforms in meters.
forward_vertices	Compute posed mesh vertices in meters.
get_rest_pose	Return zero pose controls and the model's default identity.
joint_index	Resolve a common joint to this model's native joint index.
prepare_point_regressor	Preproject a vertex mapping for repeated point forwards.
get_apose	Return the SOMA A-pose.
get_tpose	Return the SOMA T-pose.
prepare_identity	Precompute identity-dependent state for repeated forward passes.
prepare_pose	Precompute pose-dependent state for repeated forward passes.

ATTRIBUTE	DESCRIPTION
common_joints	Common anatomical joints mapped to this model's native joint names.
has_face	bool(x) -> bool
has_hands	bool(x) -> bool
num_joints	Number of joints in the skeleton.
pose_joint_indices	Canonical joints whose local transforms are driven by each pose parameter.
runtime	Array runtime used by this model.
skinning_spec	Static topology, render-rig weights, and optional pose correctives.
symmetric_joints	Left/right joint pairs as (left_index, right_index), in joint order.
NUM_BODY_CONTROLS	int([x]) -> integer
NUM_HAND_CONTROLS	int([x]) -> integer
NUM_HEAD_CONTROLS	int([x]) -> integer
NUM_JOINTS	int([x]) -> integer

common_joints property

common_joints

Common anatomical joints mapped to this model's native joint names.

has_face class-attribute

has_face = False

bool(x) -> bool

Returns True when the argument x is true, False otherwise. The builtins True and False are the only two instances of the class bool. The class bool is a subclass of the class int, and cannot be subclassed.

has_hands class-attribute

has_hands = True

bool(x) -> bool

Returns True when the argument x is true, False otherwise. The builtins True and False are the only two instances of the class bool. The class bool is a subclass of the class int, and cannot be subclassed.

num_joints property

num_joints

Number of joints in the skeleton.

pose_joint_indices property

pose_joint_indices

Canonical joints whose local transforms are driven by each pose parameter.

runtime property

runtime

Array runtime used by this model.

skinning_spec property

skinning_spec

Static topology, render-rig weights, and optional pose correctives.

symmetric_joints property

symmetric_joints

Left/right joint pairs as (left_index, right_index), in joint order.

Indices address the J axis of :meth:forward_skeleton outputs and cover the whole native skeleton, including joints outside the :class:Joint vocabulary. Unpaired joints lie on the midline. Pairs describe index correspondence only, not how to mirror a pose.

RAISES	DESCRIPTION
ValueError	If a sided joint name has no counterpart.

NUM_BODY_CONTROLS class-attribute

NUM_BODY_CONTROLS = 23

int([x]) -> integer int(x, base=10) -> integer

Convert a number or string to an integer, or return 0 if no arguments are given. If x is a number, return x.int(). For floating point numbers, this truncates towards zero.

If x is not a number or if base is given, then x must be a string, bytes, or bytearray instance representing an integer literal in the given base. The literal can be preceded by '+' or '-' and be surrounded by whitespace. The base defaults to 10. Valid bases are 0 and 2-36. Base 0 means to interpret the base from the string as an integer literal.

int('0b100', base=0) 4

NUM_HAND_CONTROLS class-attribute

NUM_HAND_CONTROLS = 48

int([x]) -> integer int(x, base=10) -> integer

Convert a number or string to an integer, or return 0 if no arguments are given. If x is a number, return x.int(). For floating point numbers, this truncates towards zero.

If x is not a number or if base is given, then x must be a string, bytes, or bytearray instance representing an integer literal in the given base. The literal can be preceded by '+' or '-' and be surrounded by whitespace. The base defaults to 10. Valid bases are 0 and 2-36. Base 0 means to interpret the base from the string as an integer literal.

int('0b100', base=0) 4

NUM_HEAD_CONTROLS class-attribute

NUM_HEAD_CONTROLS = 5

int([x]) -> integer int(x, base=10) -> integer

Convert a number or string to an integer, or return 0 if no arguments are given. If x is a number, return x.int(). For floating point numbers, this truncates towards zero.

If x is not a number or if base is given, then x must be a string, bytes, or bytearray instance representing an integer literal in the given base. The literal can be preceded by '+' or '-' and be surrounded by whitespace. The base defaults to 10. Valid bases are 0 and 2-36. Base 0 means to interpret the base from the string as an integer literal.

int('0b100', base=0) 4

NUM_JOINTS class-attribute

NUM_JOINTS = 77

int([x]) -> integer int(x, base=10) -> integer

Convert a number or string to an integer, or return 0 if no arguments are given. If x is a number, return x.int(). For floating point numbers, this truncates towards zero.

If x is not a number or if base is given, then x must be a string, bytes, or bytearray instance representing an integer literal in the given base. The literal can be preceded by '+' or '-' and be surrounded by whitespace. The base defaults to 10. Valid bases are 0 and 2-36. Base 0 means to interpret the base from the string as an integer literal.

int('0b100', base=0) 4

apply_pose_correctives

apply_pose_correctives(*, identity, pose)

Apply prepared pose correctives to identity-dependent rest vertices.

Source code in src/body_models/_base.py

def apply_pose_correctives(
    self,
    *,
    identity: SkinningIdentity,
    pose: SkinningPose,
) -> Float[Array, "*batch V 3"]:
    """Apply prepared pose correctives to identity-dependent rest vertices."""
    vertices = identity["rest_vertices"]
    coefficients = pose.get("pose_coefficients")
    if coefficients is None:
        return vertices
    basis = self._corrective_basis
    if basis is None:
        raise RuntimeError("Prepared pose has corrective coefficients, but the model has no corrective basis.")
    return vertices + basis.apply(coefficients)

forward_points

forward_points(
    body_pose,
    head_pose,
    hand_pose,
    *,
    point_regressor,
    shape=None,
    scale_params=None,
    identity=None,
    global_rotation=None,
    global_translation=None,
)

Compute positions defined by a prepared vertex mapping.

Source code in src/body_models/soma/_model.py

def forward_points(
    self,
    body_pose: Float[Array, "*batch 23 N"] | Float[Array, "*batch 23 3 3"],
    head_pose: Float[Array, "*batch 5 N"] | Float[Array, "*batch 5 3 3"],
    hand_pose: Float[Array, "*batch 48 N"] | Float[Array, "*batch 48 3 3"],
    *,
    point_regressor: PointRegressor,
    shape: Float[Array, "*batch I"] | None = None,
    scale_params: Float[Array, "*batch K"] | None = None,
    identity: SomaIdentity | None = None,
    global_rotation: Float[Array, "*batch N"] | Float[Array, "*batch 3 3"] | None = None,
    global_translation: Float[Array, "*batch 3"] | None = None,
) -> Float[Array, "*batch P 3"]:
    """Compute positions defined by a prepared vertex mapping."""
    xp = self._runtime.xp
    self._validate_identity_arguments(identity, shape=shape, scale_params=scale_params)
    batch_shape = body_pose.shape[: -(self._num_rot_dims + 1)]
    if identity is None:
        resolved = self._resolve_identity_coefficients(batch_shape, shape=shape)
        if scale_params is not None:
            scale_params = xp.broadcast_to(scale_params, (*batch_shape, scale_params.shape[-1]))
        identity = self.prepare_identity(*resolved, scale_params=scale_params)

    pose = self.prepare_pose(body_pose, head_pose, hand_pose, identity=identity)
    return self._deform_points(point_regressor, identity, pose, global_rotation, global_translation)

forward_skeleton

forward_skeleton(
    body_pose,
    head_pose,
    hand_pose,
    *,
    shape=None,
    scale_params=None,
    identity=None,
    global_rotation=None,
    global_translation=None,
    joint_indices=None,
)

Compute posed public-joint transforms in meters.

Source code in src/body_models/soma/_model.py

def forward_skeleton(
    self,
    body_pose: Float[Array, "*batch 23 N"] | Float[Array, "*batch 23 3 3"],
    head_pose: Float[Array, "*batch 5 N"] | Float[Array, "*batch 5 3 3"],
    hand_pose: Float[Array, "*batch 48 N"] | Float[Array, "*batch 48 3 3"],
    *,
    shape: Float[Array, "*batch I"] | None = None,
    scale_params: Float[Array, "*batch K"] | None = None,
    identity: SomaIdentity | None = None,
    global_rotation: Float[Array, "*batch N"] | Float[Array, "*batch 3 3"] | None = None,
    global_translation: Float[Array, "*batch 3"] | None = None,
    joint_indices: Sequence[int] | None = None,
) -> Float[Array, "*batch 77 4 4"]:
    """Compute posed public-joint transforms in meters."""
    xp = self._runtime.xp
    self._validate_identity_arguments(identity, shape=shape, scale_params=scale_params)
    batch_shape = body_pose.shape[: -(self._num_rot_dims + 1)]
    if identity is None:
        resolved = self._resolve_identity_coefficients(batch_shape, shape=shape)
        if scale_params is not None:
            scale_params = xp.broadcast_to(scale_params, (*batch_shape, scale_params.shape[-1]))
        skeleton_identity = self._prepare_skeleton_identity(*resolved, scale_params=scale_params)
    else:
        skeleton_identity = identity

    root_rotation = SO3.identity_as(
        body_pose,
        batch_dims=batch_shape,
        rotation_type=self.rotation_type,
        xp=xp,
    )
    pose = pose_utils.pack_pose(xp, root_rotation, body_pose, head_pose, hand_pose)
    skeleton = core.prepare_skeleton(
        self._runtime,
        self._assets,
        pose,
        self.rotation_type,
        local_joint_translations=skeleton_identity["local_joint_translations"],
        joint_indices=joint_indices,
    )
    return skinning.transform_skeleton(
        skeleton,
        global_rotation,
        global_translation,
        self.rotation_type,
        xp=xp,
    )

forward_vertices

forward_vertices(
    body_pose,
    head_pose,
    hand_pose,
    *,
    shape=None,
    scale_params=None,
    identity=None,
    global_rotation=None,
    global_translation=None,
    vertex_indices=None,
)

Compute posed mesh vertices in meters.

Source code in src/body_models/soma/_model.py

def forward_vertices(
    self,
    body_pose: Float[Array, "*batch 23 N"] | Float[Array, "*batch 23 3 3"],
    head_pose: Float[Array, "*batch 5 N"] | Float[Array, "*batch 5 3 3"],
    hand_pose: Float[Array, "*batch 48 N"] | Float[Array, "*batch 48 3 3"],
    *,
    shape: Float[Array, "*batch I"] | None = None,
    scale_params: Float[Array, "*batch K"] | None = None,
    identity: SomaIdentity | None = None,
    global_rotation: Float[Array, "*batch N"] | Float[Array, "*batch 3 3"] | None = None,
    global_translation: Float[Array, "*batch 3"] | None = None,
    vertex_indices: Sequence[int] | None = None,
) -> Float[Array, "*batch V 3"]:
    """Compute posed mesh vertices in meters."""
    xp = self._runtime.xp
    self._validate_identity_arguments(identity, shape=shape, scale_params=scale_params)
    batch_shape = body_pose.shape[: -(self._num_rot_dims + 1)]
    if identity is None:
        resolved = self._resolve_identity_coefficients(batch_shape, shape=shape)
        if scale_params is not None:
            scale_params = xp.broadcast_to(scale_params, (*batch_shape, scale_params.shape[-1]))
        identity = self.prepare_identity(*resolved, scale_params=scale_params)

    pose = self.prepare_pose(body_pose, head_pose, hand_pose, identity=identity)
    vertices = self._runtime._skin_vertices(
        self.apply_pose_correctives(identity=identity, pose=pose),
        pose["skinning_transforms"],
        skinning=self._assets.compact_skinning,
        vertex_indices=vertex_indices,
    )
    return skinning.apply_global_transform(
        vertices,
        global_rotation,
        global_translation,
        self.rotation_type,
        xp=xp,
    )

get_rest_pose

get_rest_pose(*, batch_dims=(), dtype=None, hands='default')

Return zero pose controls and the model's default identity.

Source code in src/body_models/soma/_model.py

def get_rest_pose(
    self,
    *,
    batch_dims: tuple[int, ...] = (),
    dtype: Any | None = None,
    hands: Literal["default", "flat", "rest"] = "default",
) -> dict[str, Float[Array, "..."]]:
    """Return zero pose controls and the model's default identity."""
    if hands not in ("default", "flat", "rest"):
        raise ValueError(f"Invalid hands: {hands!r}. Expected 'default', 'flat', or 'rest'.")

    params = super().get_rest_pose(batch_dims=batch_dims, dtype=dtype)
    if hands != "default":
        runtime = self.runtime
        axis_angle = runtime.asarray(SOMA_HAND_PRESETS[hands], like=params["hand_pose"]).reshape(-1, 3)
        axis_angle = runtime.xp.broadcast_to(axis_angle, (*batch_dims, *axis_angle.shape))
        params["hand_pose"] = SO3.convert(
            axis_angle,
            src="axis_angle",
            dst=self.rotation_type,
            xp=runtime.xp,
        )
    return params

joint_index

joint_index(joint)

Resolve a common joint to this model's native joint index.

Source code in src/body_models/_base.py

def joint_index(self, joint: Joint) -> int:
    """Resolve a common joint to this model's native joint index."""
    if not isinstance(joint, Joint):
        raise TypeError("joint_index() expects a body_models.Joint; use joint_names.index(...) for native names.")
    try:
        native_name = self.common_joints[joint]
    except KeyError as exc:
        raise KeyError(f"{self.__class__.__name__} has no common joint {joint.value!r}") from exc
    return self.joint_names.index(native_name)

prepare_point_regressor

prepare_point_regressor(mapping)

Preproject a vertex mapping for repeated point forwards.

For Torch, call this after moving the model to its target device.

Source code in src/body_models/_base.py

def prepare_point_regressor(
    self,
    mapping: Float[Array, "K V"],
) -> PointRegressor:
    """Preproject a vertex mapping for repeated point forwards.

    For Torch, call this after moving the model to its target device.
    """
    if mapping.ndim != 2 or mapping.shape[0] < 1 or mapping.shape[1] != self.num_vertices:
        raise ValueError(
            f"mapping must have shape [K, {self.num_vertices}] with K >= 1, got {tuple(mapping.shape)}"
        )
    mapping = self._runtime.asarray(mapping, like=self.rest_vertices)
    return point_regression.prepare_point_regressor(
        mapping,
        self._skinning_weights,
        self._corrective_basis,
        runtime=self._runtime,
    )

get_apose

get_apose(*, batch_dims=(), dtype=None, hands='default')

Return the SOMA A-pose.

Source code in src/body_models/soma/_model.py

def get_apose(
    self,
    *,
    batch_dims: tuple[int, ...] = (),
    dtype: Any | None = None,
    hands: Literal["default", "flat", "rest"] = "default",
) -> dict[str, Float[Array, "..."]]:
    """Return the SOMA A-pose."""
    params = self.get_rest_pose(batch_dims=batch_dims, dtype=dtype, hands=hands)
    xp = self._runtime.xp
    axis_angle = self._runtime.asarray(SOMA_BODY_PRESETS["a_pose"], like=params["body_pose"])
    axis_angle = xp.broadcast_to(axis_angle, (*batch_dims, *axis_angle.shape))
    params["body_pose"] = SO3.convert(axis_angle, src="axis_angle", dst=self.rotation_type, xp=xp)
    return params

get_tpose

get_tpose(*, batch_dims=(), dtype=None, hands='default')

Return the SOMA T-pose.

Source code in src/body_models/soma/_model.py

def get_tpose(
    self,
    *,
    batch_dims: tuple[int, ...] = (),
    dtype: Any | None = None,
    hands: Literal["default", "flat", "rest"] = "default",
) -> dict[str, Float[Array, "..."]]:
    """Return the SOMA T-pose."""
    return self.get_rest_pose(batch_dims=batch_dims, dtype=dtype, hands=hands)

prepare_identity

prepare_identity(
    shape, *, scale_params=None, repose=True, bind_pose="fit"
)

Precompute identity-dependent state for repeated forward passes.

Source code in src/body_models/soma/_model.py

def prepare_identity(
    self,
    shape: Float[Array, "*batch I"],
    *,
    scale_params: Float[Array, "*batch K"] | None = None,
    repose: bool = True,
    bind_pose: core.BindPoseMode = "fit",
) -> SomaIdentity:
    """Precompute identity-dependent state for repeated forward passes."""
    rest_shape_full, rest_shape_active = self._rest_shapes(shape, scale_params)
    return core.prepare_identity_from_rest_shape(
        self._runtime,
        data=self._assets,
        rest_shape_full=rest_shape_full,
        rest_shape_active=rest_shape_active,
        repose=repose,
        bind_pose=bind_pose,
    )

prepare_pose

prepare_pose(body_pose, head_pose, hand_pose, *, identity)

Precompute pose-dependent state for repeated forward passes.

Source code in src/body_models/soma/_model.py

def prepare_pose(
    self,
    body_pose: Float[Array, "*batch 23 N"] | Float[Array, "*batch 23 3 3"],
    head_pose: Float[Array, "*batch 5 N"] | Float[Array, "*batch 5 3 3"],
    hand_pose: Float[Array, "*batch 48 N"] | Float[Array, "*batch 48 3 3"],
    *,
    identity: SomaIdentity,
) -> SkinningPose:
    """Precompute pose-dependent state for repeated forward passes."""
    xp = self._runtime.xp
    batch_shape = body_pose.shape[: -(self._num_rot_dims + 1)]
    root_rotation = SO3.identity_as(
        body_pose,
        batch_dims=batch_shape,
        rotation_type=self.rotation_type,
        xp=xp,
    )
    pose = pose_utils.pack_pose(xp, root_rotation, body_pose, head_pose, hand_pose)
    return core.prepare_pose(
        self._runtime,
        self._assets,
        pose,
        rotation_type=self.rotation_type,
        local_joint_translations=identity["local_joint_translations"],
        inverse_bind_transforms=identity["inverse_bind_transforms"],
    )