body-models-viser

Browser-side body model evaluation for viser.

Python owns the body model asset loaders and the prepare_identity() / prepare_pose() calls. TypeScript owns the browser model lifecycle and WASM buffers. Rust owns the stateless forward_vertices() kernel.

Usage

Skinned Body Models

Use add_body_model() for non-rigid skinned body models such as SMPL, SMPL-X, MANO, FLAME, MHR, SKEL, ANNY, GarmentMeasurements, and SOMA.

import body_models_viser as bmv
import viser
from body_models.smpl.numpy import SMPL

server = viser.ViserServer()
model = SMPL(gender="neutral")
handle = bmv.add_body_model(server.scene, "/smpl", model, color=(173, 216, 230))

pose = handle.body_pose.copy()
pose[2, 0] = 0.5
handle.body_pose = pose

translation = handle.global_translation.copy()
translation[1] = 0.25
handle.set_transform(global_translation=translation)

shape = handle.shape.copy()
shape[0] = 1.0
handle.set_identity(shape=shape)

add_body_model() returns a generic handle with set_identity(...), set_pose(...), set_transform(...), remove(), global_rotation, global_translation, and the parameter properties declared by the model.

Skeletons

Use add_skeleton() for a standalone clickable skeleton. It takes joint positions and a parent index for each joint.

import body_models_viser as bmv

pose = model.get_rest_pose()
skeleton = model.forward_skeleton(**pose)
joint_positions = skeleton[:, :3, 3]

handle = bmv.add_skeleton(
    server.scene,
    "/skeleton",
    joint_positions,
    model.parents,
    joint_names=tuple(model.joint_names),
)

handle.visible = True
handle.joint_positions = joint_positions

Rigid Body Models

Use add_rigid_body_model() for any body_models.RigidBodyModel, such as G1, BrainCo, SmplHumanoid, and MyoFullBody.

import body_models_viser as bmv
from body_models import create_model

model = create_model("g1")
handle = bmv.add_rigid_body_model(server.scene, "/robot", model)
handle.set_pose(body_pose=handle.pose["body_pose"])

The rigid-body helper bakes one static link-local mesh per link from the rest pose (forward_meshes() sliced with the model's link metadata), then only updates link transforms from forward_links() when the pose changes.

Runtime

bmv.add_body_model(scene, name, model) does three things:

1. Injects body-models-viser.js and body-models-viser.wasm.
2. Sends faces, material props, current identity, and current pose as little-endian binary buffers.
3. Returns a body model handle.

handle.set_identity(...) recomputes prepare_identity(), then prepares and sends updated skinning data. handle.set_pose(...) reuses the cached identity and sends updated pose-dependent skinning data. handle.set_transform(...) updates only global transform buffers. TypeScript copies changed buffers into persistent WASM memory, calls the Rust forward_vertices() kernel, and forwards the resulting vertex buffer to viser as a regular mesh message.

The browser protocol is model-agnostic: Python converts every SkinnedModel to a shared runtime payload containing skinning weights, rest vertices, skinning transforms, and pose offsets. Models that implement the standard preparation API and parameter-key metadata require no viser-specific adapter.

viser compatibility

This package patches viser private internals (message serializer, websock client state, and the client React tree) to inject its runtime. The supported viser range is pinned in pyproject.toml; when raising the ceiling, run uv run pytest and uv run scripts/visualize_models.py against the new version and check a browser actually renders.

Development

Build the Rust crate:

cargo test

Build the browser bundle and WASM:

cd client
npm test

The browser build requires a Rust toolchain with wasm32-unknown-unknown installed, for example:

rustup target add wasm32-unknown-unknown

Run the small visualizer:

uv run --no-sync scripts/visualize_models.py

Check NumPy/WASM vertex parity for all supported models:

uv run scripts/check_model_parity.py