ISO 8559-1 body measurements for Anny and MHR parametric body models. Twelve keys are differentiable through PyTorch autograd for gradient-based body fitting.

Anny and MHR give you a 14–18K vertex mesh and nothing to measure it with. SMPL tooling doesn't port over, and the plane-sweep algorithms look simple until you hit convex-hull tape simulation, contour-fragment merging, and ISO-compliant landmark detection for bust/hip/crotch. clad-body is that work, done once — 25 anthropometric measurements over circumferences, lengths, and body composition (volume, mass, BMI, body fat), calibrated against real scan data. It's used in production at Clad for size-aware virtual try-on.

Prefer REST? Same engine is available as a hosted API at api.clad.you — questionnaire or photo → GLB mesh + the ISO measurements below. Free tier (5 calls/week), grab a key at clad.you/developers. Good fit if you want the body model without the Python install or the Anny model weights.

All bodies are normalised to the same coordinate convention: Z-up, metres, XY-centred, feet at Z=0, +Y=front.

pip install clad-body

# With Anny body loader (requires torch)
pip install 'clad-body[anny]'

# With MHR body loader (requires pymomentum)
pip install 'clad-body[mhr]'

# With 4-view rendering
pip install 'clad-body[render]'

from clad_body.load import load_anny_from_params
from clad_body.measure import measure

body = load_anny_from_params(params)

m = measure(body)                                  # all measurements
m = measure(body, preset="core")                   # 4: height, bust, waist, hip
m = measure(body, preset="standard")               # 9: + thigh, upperarm, shoulder, sleeve, inseam
m = measure(body, preset="tops")                   # garment-relevant subset
m = measure(body, only=["bust_cm", "hip_cm"])       # specific keys
m = measure(body, tags={"type": "circumference", "region": "leg"})  # tag filter
m = measure(body, render_path="body.png")          # with 4-view render

MHR works the same way:

from clad_body.load import load_mhr_from_params
body = load_mhr_from_params("path/to/sam3d_params.json")
m = measure(body)

Under active development. API surface and supported keys may change between minor versions. Twelve keys are differentiable today; more will follow.

For autograd-based optimization of the body mesh, use measure_grad(body) instead of measure(body). Same input, same key names — but the returned values are PyTorch tensors with autograd history, so you can put them directly into a loss and backprop into the Anny phenotype parameters.

Pass requires_grad=True to load_anny_from_params to create the body with gradient-enabled phenotype tensors (stored on body.phenotype_kwargs):

import torch
from clad_body.load import load_anny_from_params
from clad_body.measure import measure_grad

body = load_anny_from_params(initial_params, requires_grad=True)
optimizer = torch.optim.Adam(list(body.phenotype_kwargs.values()), lr=0.01)

for step in range(500):
    optimizer.zero_grad()
    m = measure_grad(body, only=["bust_cm", "waist_cm", "inseam_cm"])
    loss = (m["bust_cm"] - 92.0) ** 2 + (m["waist_cm"] - 78.0) ** 2 + (m["inseam_cm"] - 82.0) ** 2
    loss.backward()
    optimizer.step()

Each measure_grad(body) call re-runs the forward pass using body.phenotype_kwargs, so after optimizer.step() updates the tensors the next iteration measures the new mesh. If you only want to optimize a subset of parameters, load without requires_grad=True and enable it per-tensor: body.phenotype_kwargs["height"].requires_grad_(True).

Supported keys and their calibration error vs the ISO reference that measure() uses:

Both measure() and measure_grad report convex hull circumference, not the raw cross-section perimeter. This matches ISO 8559-1: a real measuring tape bridges across concavities (e.g., cleavage between breasts, armpit crease) rather than dipping into them. The convex hull perimeter is always ≤ the raw contour perimeter — the difference is most visible at the bust on larger cup sizes where the cleavage concavity can shorten the measurement by 1-3 cm compared to following the actual surface.

measure_grad builds a 72-point polygon via differentiable soft edge-plane intersection, then takes its scipy.spatial.ConvexHull perimeter. This is used for bust, underbust, hip, and stomach — each at its respective anatomical height. Every measure_grad() call recomputes the hull from scratch (new forward pass → new polygon → new hull). The hull decides which polygon vertices to keep (discrete, like argmax or tensor[mask]) — but the perimeter of those vertices is a plain sum of torch.linalg.norm over their positions, so loss.backward() flows gradients through the kept vertices back to the Anny phenotype params. Dropped vertices (inside the hull, e.g., cleavage or gluteal cleft bins) get zero gradient — correct, since they don't affect the tape measure. The hull indices can change between optimization steps if the body shape changes enough, but the perimeter value is continuous at those transitions so the loss doesn't jump. See clad_body/measure/_soft_circ.py.

For stomach_cm the cutting-plane height is itself differentiable: a soft-argmin over torso vertex Y (most anterior = most negative) in the belly Z-range picks the height of maximum anterior protrusion, and one soft_circumference is taken there. See measure_stomach_soft in the same file. The Z selection can drift by 1–2 cm from the reference's 2 mm band-scan argmax on bodies where multiple vertex clusters have near-identical anterior Y — which translates to ~3 cm of circumference error because the body tapers rapidly in the belly region. This is why stomach_cm has looser tolerance than the other soft-circ keys.

Strictly speaking, stomach_cm is differentiable almost everywhere, with numerically-tiny (~10⁻⁵ cm) step discontinuities where a vertex crosses the hard Z-mask boundary that gates out feet and bust from the belly soft-argmin.

Requesting any other key raises ValueError. There is no silent numpy fallback — it would break gradient flow without warning. For non-differentiable keys use measure().

Jacobian heatmap for every Anny local_changes dimension against every supported measurement — |d(measurement)/d(param)| averaged across 6 reference bodies. Darker cells mean the local_change has strong leverage over that measurement; pale cells mean the gradient is near zero and Adam will barely move it. Useful when deciding which params to unfreeze for a given fit target (e.g. optimising bust_cm is pointless without measure-bust-circ-incr or torso-scale-horiz-incr in the active set). Regenerate with python tools/sensitivity_map.py --output <path>.

measure() accepts preset, only, tags, exclude. Precedence: only > preset > tags > default ("all"). exclude is applied last. Only runs computation groups needed for the requested keys.

from clad_body.measure import REGISTRY, list_measurements

REGISTRY["bust_cm"].description   # self-measurement instructions
REGISTRY["bust_cm"].iso_ref       # "5.3.4"
REGISTRY["bust_cm"].type          # "circumference"

list_measurements(type="circumference", region="leg")   # [thigh, knee, calf]

Every measurement is tagged across 5 dimensions. Each carries a human-readable description for self-measurement instructions and i18n key mapping.

Garment codes: Tops, Bottoms, Dresses, Outerwear, Underwear.

Contour	Key	Description	ISO	Type	Std	Region	Tier	Grp	Gar
	height_cm	Vertical distance from floor to top of head. Stand erect, feet together.	5.1.1	scalar	iso	full_body	core	A	all
	bust_cm	Horizontal circumference at the fullest part of the chest/bust. Tape under armpits, across bust prominence, level and snug.	5.3.4	circ	iso	torso	core	A	T,D,O,U
	waist_cm	Horizontal circumference at natural waist, midway between lowest rib and hip bone. Tape at navel height, parallel to floor.	5.3.10	circ	iso	torso	core	A	all
	hip_cm	Horizontal circumference at greatest buttock prominence. Feet together, tape around widest part of hips.	5.3.13	circ	iso	abdomen	core	A	B,D,O,U
	thigh_cm	Horizontal circumference at fullest part of upper thigh, just below gluteal fold. Stand with legs slightly apart.	5.3.20	circ	iso	leg	std	B	B
	upperarm_cm	Circumference at fullest part of upper arm, midway between shoulder and elbow. Arm relaxed, not flexed.	5.3.16	circ	iso	arm	std	B	T,O
	shoulder_width_cm	Distance between left and right shoulder points (acromion), measured across back over C7 vertebra.	5.4.2	length	iso	torso	std	C	T,D,O
	sleeve_length_cm	Distance from shoulder point along outside of slightly bent arm, over elbow, to wrist bone. (ISO §5.4.14 + §5.4.15 outer arm length, computed via plane-slice surface walk on rest pose; differentiable runtime is bone chain + linear correction.)	5.7.8	length	iso	arm	std	C	T,O
	inseam_cm	Distance from crotch point straight down to floor. Stand erect, feet slightly apart.	5.1.15	length	iso	leg	std	E	B
	neck_cm	Circumference just below Adam's apple, perpendicular to neck axis. Comfortably snug.	5.3.2	circ	iso	neck	enh	D	T
	underbust_cm	Horizontal circumference directly below breast tissue, at inframammary crease. Bra band size.	5.3.6	circ	iso	torso	enh	A	T,D,U
	stomach_cm	Horizontal circumference at maximum anterior protrusion of abdomen, usually at/below navel.	--	circ	tailor	abdomen	enh	A	T,B
	mass_kg	Total body mass in kilograms.	5.6.1	scalar	iso	full_body	enh	G	--
	volume_m3	Total body volume in cubic metres, from mesh geometry.	--	scalar	derived	full_body	enh	G	--
	bmi	Body mass index: mass (kg) / height (m)^2.	--	scalar	derived	full_body	enh	G	--
	body_fat_pct	Estimated body fat % via Navy/Weltman equations from circumferences.	--	scalar	derived	full_body	enh	G	--
	belly_depth_cm	How much belly protrudes forward vs underbust/ribcage. Negative = belly prominence.	--	scalar	derived	abdomen	enh	A	T,B
	knee_cm	Horizontal circumference at centre of kneecap. Bend knee slightly (~45 degrees).	5.3.22	circ	iso	leg	fit	B	B
	calf_cm	Maximum horizontal circumference of the calf. Stand with legs slightly apart.	5.3.24	circ	iso	leg	fit	B	B
	wrist_cm	Circumference at wrist, at prominent bone on little finger side (ulnar styloid).	5.3.19	circ	iso	arm	fit	D	T
	crotch_length_cm	Distance from front waist centre, through crotch, to back waist centre. Follow body surface.	5.4.18	length	iso	leg	fit	E	B
	front_rise_cm	Front waist to crotch point, along front body surface. Trouser front panel length.	--	length	tailor	leg	fit	E	B
	back_rise_cm	Back waist to crotch point, along back body surface. Trouser back panel length.	--	length	tailor	leg	fit	E	B
	shirt_length_cm	Side neck point down along front body contour to crotch level. Follow chest/stomach curve.	--	length	tailor	torso	fit	F	T
	back_neck_to_waist_cm	Cervicale (C7) down centre back along body contour to waist level. Tape follows spine curvature.	5.4.5	length	iso	torso	enh	H	T,D,O

Tier codes: core, std (standard), enh (enhanced), fit (fitted). Anny-only: underbust, mass, volume, bmi, body_fat, belly_depth.

Group C and E have differentiable alternatives in measure_grad — use it for hot-loop optimization instead of calling measure() repeatedly.

measure() only runs the computation groups needed for the requested keys — use only= or preset= to skip expensive groups:

measure(body)                         # all groups — ~800 ms
measure(body, preset="core")          # group A only — ~100 ms
measure(body, only=["bust_cm"])       # group A only — ~100 ms
measure(body, only=["shoulder_width_cm"])  # groups A + C — ~200 ms

measure() accepts a device parameter (None = auto-detect CUDA):

measure(body, only=["bust_cm"], device="cuda")  # GPU forward pass
measure(body, device=None)                       # auto: CUDA if available

Without extras, only numpy, scipy, and trimesh are required.

• Consumer demo — full pipeline (questionnaire → body → virtual try-on) at clad.you/size-aware/size-me.
• REST API — same body model + these measurements behind a bearer-token API at api.clad.you. Swagger UI on the root; 5 calls/week free tier. Key management at clad.you/developers.

This library was built for Clad's size-aware virtual try-on pipeline. Read the full story: A 3D Body Scan for Nine Cents — Without SMPL.

Apache 2.0 — see LICENSE.