MANO: A 3D Hand Model

• MANO is a low-dimensional, statistically learned 3D hand model that factorizes hand shape and pose with articulated non-rigid deformations.
• It integrates with SMPL as SMPL+H, using optimization-based fitting and priors to reliably recover detailed body–hand configurations.
• Extensions to MANO address limitations by adding textures, muscle dynamics, and rigid-body simulation, enhancing applications in digital twins and reconstruction.

MANO, the hand Model with Articulated and Non-rigid def*O*rmations, is a learned, parametric 3D hand model designed to be anatomically realistic and low-dimensional, compatible with standard animation pipelines, and robust enough to recover hand pose and shape from low-quality, noisy, and occluded full-body data. It factorizes hand geometry into identity-dependent shape and pose-dependent deformation, provides a compact mapping from hand poses to pose blend shape corrections and a linear manifold of pose synergies, and was attached to SMPL to form SMPL+H for coordinated body–hand capture (Romero et al., 2022).

1. Model definition and learned parameterization

MANO is formulated as a skinned statistical hand model in the SMPL style. Its core mapping is

$$M(\boldsymbol\beta,\boldsymbol\theta) = W\big(T_P(\boldsymbol\beta,\boldsymbol\theta),\, J(\boldsymbol\beta),\, \boldsymbol\theta,\, \mathcal{W}\big),$$

with

$$T_P(\boldsymbol\beta,\boldsymbol\theta) = \bar{T} + B_S(\boldsymbol\beta) + B_P(\boldsymbol\theta).$$

Here, $\bar{T}$ is the template hand mesh in a zero pose, $B_S(\boldsymbol\beta)$ is the shape blend-shape function, $B_P(\boldsymbol\theta)$ is the pose blend-shape function, $J(\boldsymbol\beta)$ are joint locations regressed from shape, $\boldsymbol\theta$ are pose parameters, $\mathcal{W}$ are skinning weights, and $W(\cdot)$ is linear blend skinning (Romero et al., 2022).

The shape space is linear:

$$B_S(\boldsymbol\beta;\mathcal{S}) = \sum_{n=1}^{|\mathcal{S}|} \beta_n\, S_n,$$

while the pose-corrective term is also linear in rotation-matrix deviations:

$$T_P(\boldsymbol\beta,\boldsymbol\theta) = \bar{T} + B_S(\boldsymbol\beta) + B_P(\boldsymbol\theta).$$0

This combination lets MANO preserve the computational convenience of linear blend skinning while correcting characteristic LBS artifacts around joints. The model uses 15 joints plus a global orientation for the hand, with joints modeled as 3-DOF ball joints for simplicity (Romero et al., 2022).

A defining property of MANO is its low-dimensional pose subspace. After PCA on axis-angle pose vectors, 6 components per hand capture about 81% of pose variance, 10 components about 90%, and 15 components about 95%. This low-dimensional structure is the model’s explicit encoding of pose synergies, and it is central to the model’s robustness under occlusion and low-resolution capture. The original model was learned from around 1000 high-resolution 3D scans of hands of 31 subjects in a wide variety of hand poses, and the resulting representation was presented as realistic, low-dimensional, compatible with standard graphics packages, and able to fit any human hand (Romero et al., 2022).

2. Fitting, registration, and attachment to the body

MANO was introduced in a setting where hands and bodies had usually been modeled separately, despite the fact that humans move their hands and bodies together. The model was therefore designed not only as a standalone hand representation but also as a hand module that can be attached to SMPL, yielding SMPL+H (Romero et al., 2022).

In SMPL+H, the shape space remains the full SMPL shape space, preserving correlations between body and hand morphology. The wrist joint comes from SMPL, while the finger joints, their blend weights, and the pose blend shapes come from MANO. The resulting body-and-hand model has 78 pose degrees of freedom: SMPL without hands has 66 DoF, and SMPL+H adds 6 latent pose parameters per hand. This attachment strategy preserves a single articulated body–hand skeleton while retaining detailed finger articulation (Romero et al., 2022).

The fitting pipeline is explicitly optimization-based. For scan registration, the model uses a robust scan-to-mesh term, a coupling term that keeps a free registration close to a valid model instance, and priors over pose and shape. In the first-frame body fitting stage, the hand prior is a Gaussian mixture model over hand pose with $$T_P(\boldsymbol\beta,\boldsymbol\theta) = \bar{T} + B_S(\boldsymbol\beta) + B_P(\boldsymbol\theta).$$1 mixture components per hand. For sequence capture, the pipeline first builds a subject-specific template from initial frames and then tracks the sequence with that template, using robust geometry terms together with temporal regularization and the hand-pose prior. This architecture reflects the original motivation: when scanning or capturing the full body in 3D, hands are small and often partially occluded, making their shape and pose hard to recover; MANO provides the strong prior needed to fill in plausible detail from degraded evidence (Romero et al., 2022).

3. Dataset infrastructure and representational extensions

Subsequent work has used MANO as an annotation engine, an interoperability layer, and a base geometry to be extended. In "Dense Hand-Object(HO) GraspNet with Full Grasping Taxonomy and Dynamics," accurate hand and object 3D meshes are obtained by fitting the hand parametric model (MANO) and the hand implicit function (HALO) to multi-view RGBD frames, with the MoCap system only for objects. The dataset includes 22 rigid objects from the YCB dataset and 8 other compound objects, diverse hand shapes from 99 participants aged 10 to 74, continuous video frames, and a 1.5M RGB-Depth of sparse frames with annotations; it also reports that HALO fitting does not require any parameter tuning and has comparable accuracy to MANO (Cho et al., 2024).

DART extends MANO in a different direction: appearance. The paper begins from the observation that MANO disregards textures and accessories, which largely limits its power to synthesize photorealistic hand data. DART therefore augments MANO with a wrist-enhanced template, increasing the geometry from 778 vertices and 1538 faces to 842 vertices and 1666 faces, and adds 325 hand-crafted 2D texture maps together with 50 daily 3D accessories consisting of 10 rings, 10 watches, 20 bracelets, and 10 gloves. DARTset then scales this extension into a large synthetic corpus of 800K high-fidelity hand images paired with perfect-aligned 3D labels (Gao et al., 2022).

These two lines of work highlight complementary limitations of the original model. HOGraspNet uses MANO because it is a tractable parametric prior for dense annotation and benchmarking, whereas DART modifies MANO because the original public template is wristless and carries no explicit texture or accessory model. A plausible implication is that MANO’s enduring value lies less in being exhaustive than in being a stable geometric and kinematic substrate onto which domain-specific augmentations can be attached.

4. Estimation, reconstruction, and biomechanical refinement

In reconstruction, MANO functions simultaneously as an output space, a prior, and a supervisory target. "MeMaHand: Exploiting Mesh-Mano Interaction for Single Image Two-Hand Reconstruction" predicts both full meshes and MANO parameters for left and right hands from a single RGB image. In that formulation, MANO pose parameters are $$T_P(\boldsymbol\beta,\boldsymbol\theta) = \bar{T} + B_S(\boldsymbol\beta) + B_P(\boldsymbol\theta).$$2, shape parameters are $$T_P(\boldsymbol\beta,\boldsymbol\theta) = \bar{T} + B_S(\boldsymbol\beta) + B_P(\boldsymbol\theta).$$3, and the MANO layer maps them to a triangulated mesh with 778 vertices. The paper’s Mesh-Mano interaction blocks use mesh vertices and MANO parameters as two kinds of query tokens, with asymmetric attention masks for intra-hand and inter-hand reasoning; on InterHand2.6M, the full model reports MPJPE 8.65 mm and MPVPE 8.89 mm, improving over IntagHand’s 8.79 mm and 9.03 mm, and its MANO branch also improves substantially over earlier parametric methods (Wang et al., 2023).

MANO’s original formulation is purely kinematic, and this has motivated biomechanical extensions. "MS-MANO: Enabling Hand Pose Tracking with Biomechanical Constraints" explicitly states that MANO does not include muscles, tendons, or physical dynamics, and introduces a musculoskeletal system with Hill-type muscle models to drive the MANO skeleton. The resulting MS-MANO model computes muscle torques and uses a simulation-in-the-loop refinement module, BioPR, to correct predicted pose sequences. On DexYCB, BioPR improves gSDF from MPJPE 14.4 to 12.81 and AUC 89.1 to 89.7, and improves Deformer from MPJPE 13.64 to 12.92 and AUC 89.6 to 90.4, while also slightly reducing acceleration error (Xie et al., 2024).

Taken together, these systems show that later work rarely treats MANO as sufficient by itself. Instead, it is embedded inside architectures that either compensate for its limited expressiveness with non-parametric mesh branches, or compensate for its lack of physiology with explicit muscle-driven dynamics.

5. Generative modeling, digital twins, and simulation

Recent work has also repurposed MANO outside classical reconstruction. "HandDreamer: Zero-Shot Text to 3D Hand Model Generation using Corrective Hand Shape Guidance" uses MANO as the central structural prior for zero-shot text-to-3D hand synthesis. MANO provides mesh initialization for NeRF density, hand keypoints and a 21-by-3 skeleton for ControlNet conditioning, and silhouettes for the Corrective Hand Shape loss. The CHS schedule uses $$T_P(\boldsymbol\beta,\boldsymbol\theta) = \bar{T} + B_S(\boldsymbol\beta) + B_P(\boldsymbol\theta).$$4, $$T_P(\boldsymbol\beta,\boldsymbol\theta) = \bar{T} + B_S(\boldsymbol\beta) + B_P(\boldsymbol\theta).$$5, $$T_P(\boldsymbol\beta,\boldsymbol\theta) = \bar{T} + B_S(\boldsymbol\beta) + B_P(\boldsymbol\theta).$$6, and $$T_P(\boldsymbol\beta,\boldsymbol\theta) = \bar{T} + B_S(\boldsymbol\beta) + B_P(\boldsymbol\theta).$$7, and the final method reaches a CLIP L14 score of 28.63, compared with 28.48 for Skeleton-CN plus MANO initialization and 28.02 for Skeleton-CN plus CHS without MANO initialization. The output is not a MANO mesh but a higher-resolution NeRF-derived mesh, later exported at approximately 300k vertices (Rosh et al., 6 Apr 2026).

In digital-twin and robotics contexts, MANO has been pushed in the opposite direction: from implicit or skinned geometry toward rigid-body simulation. "Multi-Rigid-Body Approximation of Human Hands with Application to Digital Twin" starts from a personalized MANO model, estimated from RGB images with HaMeR and from a 22-camera infrared motion capture system with 22 reflective markers at 120 Hz, and converts that model to a URDF hand with anatomically consistent joint axes. The rigid-body model has 20 scalar joint coordinates instead of MANO’s 45 rotational DoFs, and the key technical step is projecting MANO’s unconstrained $$T_P(\boldsymbol\beta,\boldsymbol\theta) = \bar{T} + B_S(\boldsymbol\beta) + B_P(\boldsymbol\theta).$$8 joint rotations onto this constrained kinematic space using closed-form projection for 1-DoF joints and a BCH-corrected iterative method for 2-DoF joints. The reported projection time is 0.41 ms per pose, digital-twin tracking error is 0.85 cm, and grasp success across five manipulation primitives is 77.9% (Zhao et al., 8 Dec 2025).

These two uses pull MANO in different directions—one toward generative priors for diffusion and implicit surfaces, the other toward physically simulated URDF hands—yet both rely on the same property: MANO provides a compact, anatomically plausible hand scaffold that can be interfaced with optimization, rendering, and control.

6. Limitations, scope, and disambiguation

Several limitations recur across the literature. MANO is geometric and articulated, but it does not include textures or accessories; DART addresses this by adding a wrist, UV texture maps, and accessories (Gao et al., 2022). MANO is differentiable and low-dimensional, but it is kinematic rather than physiological; MS-MANO adds a muscle–tendon network and a simulation-in-the-loop refinement framework to impose biomechanically realistic constraints (Xie et al., 2024). MANO is expressive enough for pose-and-shape estimation, but it is not directly physics-ready; the digital-twin work converts it into a URDF-compatible multi-rigid-body approximation with constrained joint axes and explicit rigid-body dynamics (Zhao et al., 8 Dec 2025). These extensions do not replace the original model so much as expose its design boundary.

A common misconception in bibliographic search is that unqualified uses of “MANO” or “Mano” always denote the hand model. In contemporary arXiv literature, the same string names unrelated systems: “MANO: Exploiting Matrix Norm for Unsupervised Accuracy Estimation Under Distribution Shifts” (Xie et al., 2024), “5G Network Slice Isolation with WireGuard and Open Source MANO” (Haga et al., 2020), the benchmarking study “On the Challenges and KPIs for Benchmarking Open-Source NFV MANO Systems: OSM vs ONAP” (Yilma et al., 2019), the 6G orchestration framework “KB-MANO” (Shokrnezhad et al., 2024), the optimizer “Mano: Restriking Manifold Optimization for LLM Training” (Gu et al., 30 Jan 2026), the GUI agent “Mano Report” (Fu et al., 22 Sep 2025), and the “Multipole Attention Neural Operator (MANO)” (Colagrande et al., 3 Jul 2025). Within computer vision and graphics, however, MANO refers specifically to the hand model with articulated and non-rigid deformations introduced as a low-dimensional, skinned, statistically learned hand representation and later integrated into SMPL+H (Romero et al., 2022).