Open-source platform for 3D AI agents. Turn text or a photo into a rigged, animated GLB avatar, give it an LLM brain, memory and a wallet, and embed it anywhere with one web component. Remote MCP server, x402 payments, Three.js, browser-native.
JavaScript 3d-generationai-agentsanimationavatardigital-humangenerative-aiglbgltfimage-to-3dllmmcpmcp-servermodel-context-protocolsolanatext-to-3dthreejsweb-componentswebglwebxrx402
"""
Isolated test for blendshapes.py: template loading, head masking, and delta
transfer onto a synthetic humanoid head. No GPU / torch / GCP:
`python3 test_blendshapes.py`.
Uses a fabricated template (a small sphere with a known "jawOpen" delta) and a
synthetic target whose head is a displaced, rescaled copy of that sphere, so
the expected transferred deltas are known analytically.
"""
import os
import sys
import tempfile
import numpy as np
from blendshapes import head_mask_from_weights, load_template, transfer_blendshapes
def _check(cond, msg):
if not cond:
print(f" FAIL {msg}")
raise AssertionError(msg)
print(f" ok {msg}")
def _sphere(n=400, radius=1.0, seed=7):
rng = np.random.default_rng(seed)
v = rng.normal(size=(n, 3))
v /= np.linalg.norm(v, axis=1, keepdims=True)
return (v * radius).astype(np.float32)
def _write_template(path):
verts = _sphere()
verts -= verts.mean(axis=0)
# One real shape (downward jaw delta on the lower hemisphere) and one
# zero shape (like tongueOut when ICT lacks a source mesh).
jaw = np.zeros_like(verts)
lower = verts[:, 1] < 0
jaw[lower, 1] = -0.1
deltas = np.stack([jaw, np.zeros_like(verts)], axis=0)
np.savez_compressed(path, names=np.array(["jawOpen", "tongueOut"]),
verts=verts, deltas=deltas)
return verts, deltas
def test_load_template():
print("test_load_template")
with tempfile.TemporaryDirectory() as tmp:
path = os.path.join(tmp, "tpl.npz")
_write_template(path)
tpl = load_template(path)
_check(tpl["names"] == ["jawOpen", "tongueOut"], "names round-trip")
_check(tpl["verts"].shape[1] == 3 and tpl["deltas"].shape[0] == 2,
"shapes round-trip")
# A malformed template (delta/vert mismatch) must fail loudly.
bad = os.path.join(tmp, "bad.npz")
np.savez_compressed(bad, names=np.array(["a"]),
verts=np.zeros((10, 3), np.float32),
deltas=np.zeros((1, 9, 3), np.float32))
try:
load_template(bad)
except RuntimeError:
_check(True, "malformed template raises")
else:
_check(False, "expected RuntimeError for malformed template")
def test_head_mask():
print("test_head_mask")
names = ["mixamorig:Hips", "mixamorig:Spine", "mixamorig:Head",
"mixamorig:LeftEye"]
w = np.zeros((5, 4), np.float32)
w[0, 0] = 1.0 # hips vertex
w[1, 2] = 0.9 # head vertex
w[2, 2] = 0.2 # weak head influence: below threshold
w[3, 3] = 0.8 # eye vertex counts as head
w[4, 1] = 1.0 # spine vertex
mask = head_mask_from_weights(w, names)
_check(mask.tolist() == [False, True, False, True, False],
f"head mask selects head+eye verts only ({mask.tolist()})")
try:
head_mask_from_weights(w, ["a", "b", "c", "d"])
except RuntimeError:
_check(True, "missing head bone raises")
else:
_check(False, "expected RuntimeError when no head bone present")
def test_transfer_geometry():
print("test_transfer_geometry")
with tempfile.TemporaryDirectory() as tmp:
path = os.path.join(tmp, "tpl.npz")
tpl_verts, tpl_deltas = _write_template(path)
tpl = load_template(path)
# Target: body verts far below, head = the template sphere scaled 2x
# and moved up to y ~ 3.
scale, offset = 2.0, np.array([0.0, 3.0, 0.0], np.float32)
head = tpl_verts * scale + offset
body = _sphere(300, radius=0.8, seed=11) + np.array([0, 0.8, 0], np.float32)
target = np.concatenate([body, head], axis=0)
mask = np.zeros(len(target), dtype=bool)
mask[len(body):] = True
out = transfer_blendshapes(tpl, target, mask)
deltas = out["deltas"]
_check(deltas.shape == (2, len(target), 3), "output covers all verts")
_check(np.abs(deltas[:, :len(body), :]).max() == 0.0,
"body verts receive zero delta")
# Head verts correspond 1:1 with template verts (exact overlap after
# alignment), so the transferred jaw delta is the template's, scaled.
head_deltas = deltas[0, len(body):, :]
expected = tpl_deltas[0] * scale
err = np.abs(head_deltas - expected).max()
_check(err < 1e-3, f"jaw deltas transferred and scaled (max err {err:.2e})")
_check(np.abs(deltas[1]).max() == 0.0, "zero template shape stays zero")
def test_transfer_falloff():
print("test_transfer_falloff")
with tempfile.TemporaryDirectory() as tmp:
path = os.path.join(tmp, "tpl.npz")
_write_template(path)
tpl = load_template(path)
# Head region includes an outlier "hair" vertex far off the template
# surface; falloff must zero it.
head = tpl["verts"].copy()
hair = np.array([[0.0, 5.0, 0.0]], np.float32)
target = np.concatenate([head, hair], axis=0)
mask = np.ones(len(target), dtype=bool)
out = transfer_blendshapes(tpl, target, mask)
_check(np.abs(out["deltas"][0, -1, :]).max() == 0.0,
"far-off vertex receives zero delta (falloff)")
def test_tiny_head_skipped():
print("test_tiny_head_skipped")
with tempfile.TemporaryDirectory() as tmp:
path = os.path.join(tmp, "tpl.npz")
_write_template(path)
tpl = load_template(path)
target = _sphere(100, seed=3)
mask = np.zeros(len(target), dtype=bool)
mask[:4] = True # too few head verts to align against
out = transfer_blendshapes(tpl, target, mask)
_check(np.abs(out["deltas"]).max() == 0.0,
"degenerate head yields all-zero deltas instead of garbage")
if __name__ == "__main__":
try:
test_load_template()
test_head_mask()
test_transfer_geometry()
test_transfer_falloff()
test_tiny_head_skipped()
except AssertionError:
print("\nFAILED")
sys.exit(1)
print("\nALL PASSED")