Anny: Open-Source Parametric Human Body Model

• Anny is a scan-free, fully differentiable parametric human body model defined by interpretable semantic phenotype axes.
• It leverages artist-authored blendshapes calibrated to WHO anthropometric data to accurately represent a diverse population.
• Its design supports efficient scan registration, synthetic data generation, and interoperability with SMPL-X and HumGen3D pipelines.

Anny is an open-source, fully differentiable, scan-free parametric human body model that offers a continuous and interpretable shape space defined by semantic phenotype axes rather than learned principal components. Its construction is grounded in artist-authored blendshapes from the MakeHuman community and is calibrated to population-level anthropometric statistics (WHO growth curves). The model’s explicit semantic control allows accurate modeling of a broad demographic span—across gender, age (from infants to elders), height, weight, body types, and specific local proportions—while avoiding both the demographic biases and proprietary restrictions of traditional scan-driven shape spaces. Anny’s computational structure supports efficient scan registration, human mesh recovery (HMR), controlled synthetic data generation, and plug-and-play interoperability with SMPL-X and HumGen3D pipelines, all without reliance on costly 3D scan datasets (Brégier et al., 5 Nov 2025).

1. Semantic Shape Space Construction

Anny defines its body shape space using a vector of $K$ phenotype parameters $p \in [0,1]^K$. Each $p_k$ encodes a normalized “axis” of human variation with direct semantic meaning—for example, gender, age, height, weight, muscle mass, head fat, pregnancy (belly protuberance), and limb proportions. The number and definitions of these axes are determined by the MakeHuman library, which provides the prototypical meshes for each phenotype.

For each axis $k$, Anny includes $M_k$ exemplars $S_{k,i}$ corresponding to manually sculpted body shapes at discrete values $t_i \in [0,1]$ along that axis (e.g., $t_0$: baby, $t_1$: toddler, ... for age). An arbitrary parameter $p$ is mapped to a body mesh by independent, piecewise-linear interpolation on each axis: $$V(p) = V_0 + \sum_{k=1}^K \left[ \sum_{i=1}^{M_k-1} w_{k,i}(p_k) (S_{k,i+1} - S_{k,i}) \right],$$ where the weights $w_{k,i}(p_k)$ ensure linear interpolation within the appropriate interval $[t_i, t_{i+1}]$, and $V_0$ is the neutral mesh. All interpolants share the same mesh topology, enforcing consistency and avoiding degenerate geometry. Unlike PCA-based models, no regularization or dimensionality reduction is required—the axes are interpretable by construction (Brégier et al., 5 Nov 2025).

2. Anthropometric Calibration and Population Grounding

To ensure demographic realism, Anny links its phenotype axes to real-world anthropometric statistics via calibration to WHO data. For each gender $g$ and age $a$, the “height” and “weight” parameters are modeled as Beta-distributed random variables ($p_\text{h}$ for height; $p_\text{u}$ for upstream weight), with

$$p_\mathrm{h} \sim \mathrm{Beta}(\alpha_h(a, g), \beta_h(a, g)), \quad p_\mathrm{u} \sim \mathrm{Beta}(\alpha_u(a, g), \beta_u(a, g)).$$

Mappings $f_h$ and $f_w$ convert these to centimeters and kilograms through monotonic transforms. The shape parameters of these Beta distributions are chosen so that means and variances exactly match published WHO growth curves—both for children and adults, and preserving height–weight–BMI relationships. This population-level grounding avoids demographic overfitting and allows random sampling that yields anthropometrically plausible meshes over the entire human lifespan (Brégier et al., 5 Nov 2025).

3. Differentiable Kinematics, Skinning, and Scan Fitting

Anny’s full forward kinematic pipeline is implemented in PyTorch and NVIDIA Warp for exact differentiation with respect to $p$ (shape) and $\theta$ (pose parameters, e.g., axis-angle or quaternion joint rotations). Given skinning weights $W_{vb}$ and bone transforms $T_b(\theta)$, the final deformed mesh is

$$V(p, \theta) = \sum_{b=1}^B W_{vb}\left(T_b(\theta) V(p)_v\right).$$

This structure enables differentiable fitting to 3D scans or point clouds via minimization: $$\mathcal{L}(p, \theta) = \frac{1}{|X|}\sum_j \min_v \|x_j - V_v(p,\theta)\|^2 + \lambda_\mathrm{pose} \|\theta-\theta_0\|^2 + \lambda_\mathrm{shape} \|p-p_0\|^2.$$ Empirically, mean point-to-mesh errors are $\sim$2.4 mm (adults, 3DBodyTex) and $\sim$2.7 mm (children), with convergence achieved in seconds and no requirement for manual landmarks (Brégier et al., 5 Nov 2025).

4. Controlled Synthetic Data Generation: The Anny-One Dataset

Leveraging calibrated sampling and differentiable mesh synthesis, Anny is used to generate Anny-One, a dataset of 800,000 photorealistic synthetic humans. Each sample is produced by: sampling $p$ from the Beta-calibrated distribution at target age/gender, sampling full-body pose from AMASS, and hand pose from GRAB. The mesh can be mapped via a learned regressor to HumGen3D format, dressed with sampled clothing, and rendered in a physically realistic scene with randomized camera, lighting, and pose conditions. The dataset provides exact ground truth for all 3D/2D mesh, keypoints, and phenotype parameters (Brégier et al., 5 Nov 2025).

5. Quantitative Evaluation and Benchmarks

Anny’s accuracy and generalization are documented on established benchmarks:

• Scan Registration: 2.4 mm mean error (adults, 3DBodyTex), 2.7 mm (child scans).
• HMR (Human Mesh Recovery): On 3DPW/EHF with HMR2.0 (ViT) and real BEDLAM training only, SMPL-X achieves MPJPE 86.0 mm and PA-MPJPE 52.0 mm, while Anny achieves 86.5/49.4 mm. On AGORA (adults plus kids), pretraining with Anny-One and finetuning achieves PA-MPJPE 48.2 mm (41.5 mm for kids), compared to 50.3 (45.6 for kids) with SMPL-X. On 3DPW/EMDB/Hi4D/CMU-Toddler, SOTA multi-person PA-MPJPE is 41.8 mm with Anny, versus 46.9 mm for the prior best.
• Interoperability: Learned mappings enable plug-and-play conversion of Anny meshes to SMPL-X or HumGen3D format, allowing use in existing pipelines (Brégier et al., 5 Nov 2025).

6. Model Release and Practical Use

Anny is distributed under the Apache 2.0 license and is implemented in Python (PyTorch), integrated with NVIDIA Warp for rapid mesh evaluations. A Blender-based synthesis pipeline supports data generation. The model may be loaded and used for sampling new body shapes or fitting to point clouds via a documented API:

from anny import AnnyBody body = AnnyBody.from_pretrained('anny_base') p = body.sample_phenotypes(age_years=25, gender='M') V = body.get_vertices(p, pose=body.rest_pose()) p_opt,θ_opt = body.fit_to_point_cloud(X_scan, init=(p,θ), lr=1e-2, iters=500) mesh = body.get_mesh(p_opt,θ_opt)

Sparse regressors for interoperability to SMPL-X or HumGen3D are provided. The model's piecewise-linear blendshape construction ensures interpretability, determinism, and extensibility for new axes (body modifications, medical pathology, etc.) (Brégier et al., 5 Nov 2025).

7. Discussion: Advantages and Limitations

Anny’s principal advantages are its:

• Semantic transparency—each axis directly corresponds to an easily understood human attribute.
• Demographic range—from infants to elders, covering realistic body types across populations.
• Absence of scan-data restrictions—no proprietary learned shape space or demographic bias.
• High-accuracy scan fitting—millimeter-level errors on established real-scan datasets.
• Flexibility for synthetic data generation—rapid, realistic population sampling.

Limitations include:

• All shape variation derives from the MakeHuman artist library; anatomical outliers or certain rare morphologies may be underrepresented.
• No high-frequency local detail (e.g., wrinkles) unless augmented downstream; the model provides the global surface and pose, not fine appearance or cloth detail.
• Expressivity is determined by the number and definition of phenotype axes and prototype shapes.

Anny demonstrates that interpretable, scan-free parametric body modeling can achieve, and in several benchmarks surpass, the performance of traditional scan-based latent PCA models while providing unique advantages in transparency, licensing, and demographic coverage (Brégier et al., 5 Nov 2025).