SMPL-X Mannequin

A rigid mannequin driven directly by SMPL-X parameters — no skinning, no blend shapes. Every part, down to the fingers, is a rigid mesh owned by one SMPL-X joint; shape coefficients change bone lengths without changing part thickness. Numpy-only and fully standalone: the SMPL-X shape response is baked into a small bundled table, so neither body-models nor the SMPL-X model files are needed.

pip install mannequin-x

Usage

from mannequin import SmplxMannequin

model = SmplxMannequin(lod=1)
params = model.get_rest_pose()
vertices = model.forward_vertices(**params)

forward_vertices, forward_skeleton, forward_links, and forward_meshes take SMPL-X parameters (axis-angle, optionally batched) and use the native SMPL-X origin: identical parameters place the pelvis and neutral foot surface in the same coordinates as SMPL-X. For repeated calls with the same betas, prepare the identity once with prepare_identity and pass it as identity=.

Viser

Each link mesh is uploaded to the viser scene once, parented to its joint's frame; pose and shape updates only send the per-joint frame transforms (~10 KB per full pose).

import viser
from mannequin import SmplxMannequin, add_mannequin

server = viser.ViserServer()
model = SmplxMannequin(lod=1)
handle = add_mannequin(server.scene, "/mannequin", model)
handle.set_shape(betas)
handle.set_pose(**model.get_apose())

set_pose accepts any subset of the SMPL-X pose parameters, and full per-frame parameter dicts work directly (shape is routed to set_shape, expression is ignored). Unchanged values are skipped.

add_mannequin takes a palette — one of the built-in armor/joint pairs in PALETTES (sand, ivory, charcoal, sage, clay, slate; see renders/palettes.jpg) or a custom (armor, joint) RGB pair.

three.js

authoring/export_glb.py exports a rigged GLB (one per LOD, attached to GitHub releases): joint nodes nested in the skeleton hierarchy with meshes parented underneath, so posing is plain local quaternions on named nodes — no skinning. The SMPL-X parameter-to-joint mapping ships in the root node's glTF extras (userData.smplx_order after loading). See examples/threejs.html for a complete animated example.

Assets

Three exactly mirrored levels of detail are bundled: lod=0 (~40k vertices), lod=1 (~15k), and lod=2 (<5k). Lower-leg length is calibrated against the SMPL-X sole vertices, so shaped identities keep the same ground plane. The editable source is authoring/mannequin.blend; the installed package ships only the compact numpy assets.