Clothed Human Layering Methods

• Clothed human layering is a representation that models a person as multiple overlapping layers, including body, visible garments, and occluded garment details.
• The approach employs methods like multi-head networks and geometric factorization to separate and jointly learn inter-dependent semantic layers.
• Layered representations have been shown to improve segmentation accuracy and support advanced applications in avatar construction, motion capture, and animation.

Clothed human layering denotes a family of representations in which a clothed person is modeled not as a single undifferentiated surface, but as multiple semantic or geometric layers whose relationships may overlap, occlude, or interact. In its most explicit recent formulation, introduced for point-cloud segmentation, a single 3D point can belong simultaneously to body, visible garment, and hidden garment layers; this replaces the standard assumption that segmentation yields disjoint sets and makes the task a vector-valued point classification problem rather than a single-label assignment (Garavaso et al., 7 Aug 2025). Across adjacent literature, the same layered viewpoint appears in avatar construction, reconstruction, rendering, and animation, where body and garments are separated into distinct meshes, fields, or Gaussian primitives to support semantic reconstruction, collision handling, editability, and cloth-specific dynamics (Xiang et al., 2021).

1. Conceptual definition and scope

The core conceptual distinction in clothed human layering is between mutually exclusive labeling and overlapping membership. Standard 3D segmentation methods are designed for scene understanding and interpretation, and in that setting each point is assigned exactly one label. For clothed bodies, this is too restrictive: a visible surface point may belong to an outer garment while also determining which body region lies underneath and which garment region is hidden below it. The canonical example given in the layered-segmentation literature is a person wearing a t-shirt over trousers, where a point near the waist can simultaneously belong to the body, the trousers, and the t-shirt layer (Garavaso et al., 7 Aug 2025).

This layered interpretation appears in several forms. In semantic segmentation, it means that the output is a multidimensional label vector with one component per layer. In geometry reconstruction, it means separate recovery of an inner body and one or more outer garments. In avatar systems, it means separate latent spaces, templates, or primitives for body and clothing. In rendering systems, it means that depth ordering and occlusion are handled by explicit composition rules rather than by absorbing cloth into a single body proxy. A two-layer mesh representation with an inner body layer and an upper-clothing layer was used for animatable avatars from multi-view video, precisely because a single-layer surface forces body motion and loose garment motion into one mesh and produces ghosting and blurry boundaries (Xiang et al., 2021).

A related but broader formulation appears in monocular performance capture, where the clothed human is decomposed into a body mesh layer plus garment piece layers for geometry reconstruction, and into shading and albedo layers for texture rendering. In that setting, layering is both semantic and physical: the body serves as an articulated scaffold, garments are reconstructed as separate cloth elements, and appearance is decomposed so that wrinkles, lighting, and intrinsic color are not conflated (Su et al., 2020).

2. Representation regimes

The literature contains three recurrent representational regimes.

The first regime is explicit overlapping semantics. In the point-cloud formulation of clothed human layering, a backbone such as PointNet++, DGCNN, or Point Transformer extracts per-point features, and multiple MLP heads perform layer-specific classification from the same embedding. Each layer is treated as a separate per-point classification problem, but the predictions are learned jointly within one architecture. This differs from a trivial baseline in which each layer would be learned by an independent network; the unified multi-head design is described as more efficient and better able to capture inter-layer correlations (Garavaso et al., 7 Aug 2025).

The second regime is explicit geometric factorization. ReMu reconstructs multi-layer clothed humans from “Image Layers,” in which a subject is photographed sequentially while wearing different clothing layers with a single RGB camera. Each garment layer is reconstructed and then aligned into a shared canonical coordinate system defined by a canonical body pose using inverse linear blend skinning. Garments are represented in a template-free and category-agnostic way by unsigned distance fields, and a collision-aware optimization processes layers from inner to outer so that vertices belonging to garment regions remain just outside the preceding layer when penetration occurs (Vuran et al., 2 Aug 2025). Closely related explicit factorizations appear in layer-wise NeRF systems, tri-plane diffusion models, and Gaussian-splatting avatars, where body and clothes are represented by separate networks or primitive sets rather than one fused field (Hu et al., 2023).

The third regime is implicit layering through body priors. HiLo reconstructs a single clothed human surface from an RGB image while using a parametric naked-body model as geometric and semantic guidance. It does not explicitly separate body and garments; instead, the body prior supplies high-frequency signed-distance cues for detail and low-frequency voxel cues for robustness, and the network predicts the occupied clothed surface (Yang et al., 2024). PIFu is similarly well suited to layered clothing, arbitrary topology, and unseen regions because its pixel-aligned implicit function can represent clothing and hair without a fixed body template, but it still outputs a single clothed surface rather than an explicit layer decomposition (Saito et al., 2019).

3. Semantic segmentation and layered labeling

Layer-aware segmentation has been formalized most directly for 3D scans. The synthetic dataset created for clothed human layering starts from synthetic clothed human meshes derived from CLOTH3D/GIM3D-plus garments, combines garments into full outfits, augments data by sampling up to three animation frames per subject, and passes the meshes through a structured-light scanner simulator. Each mesh is scanned from 13 viewpoints chosen to avoid capturing the underside of the feet, and the resulting dataset contains 3306 labeled meshes with about 50,000 vertices on average. Because the synthetic garments remain available separately, overlapping regions can be labeled as belonging to both visible and occluded garments, while body labels are inferred by projecting SMPL vertices onto the scan (Garavaso et al., 7 Aug 2025).

The segmentation study evaluates five strategies. Strategy 1 is a standard single-head setup with an added overlap label but no explicit body layer. Strategy 2 performs multi-layer segmentation with implicit overlap across body, upper garment, and lower garment layers. Strategy 3 makes overlap explicit by dedicating one layer to visible clothing and another to hidden clothing. Strategy 4 refines Strategy 2 with fine-grained garment classes for upper and lower body. Strategy 5 refines Strategy 3 by classifying both visible and hidden layers into fine-grained garment classes. The body itself is estimated via the SMPL template as a source of correspondence, and under-cloth body regions are inferred from the “tightest garment parts” (Garavaso et al., 7 Aug 2025).

The empirical trends are consistent. Point Transformer performs best across all strategies, outperforming PointNet++ and DGCNN. Data augmentation helps Point Transformer and PointNet++, but does not reliably improve DGCNN. Coarse segmentation is easier than fine-grained segmentation. Explicit overlap generally performs better than implicit overlap because the label space is less ambiguous. Hidden-layer prediction is the hardest case, especially when the network must infer occluded garment parts rather than classify visible surfaces. On the synthetic dataset, the best Point Transformer results with augmentation are about 92.1 mIoU for Strategy 1, around 89.0 average mIoU for Strategy 2, around 86.2 for Strategy 3, about 82.8 for Strategy 4, and about 79.0 for Strategy 5; qualitative transfer to real CloSe scans is reported as visually consistent despite the absence of layer-wise ground truth (Garavaso et al., 7 Aug 2025).

4. Reconstruction from images, scans, and video

Layered reconstruction pipelines typically begin by canonicalizing observations and then enforcing inter-layer consistency. ReMu is a representative example. For each image layer, it fits an SMPL-X body model, reconstructs a single-layer watertight clothed mesh with an off-the-shelf single-view human reconstruction model, isolates the target garment by rendering 60 circular views and segmenting them with Grounded SAM, and reprojects the resulting masks back onto the mesh. After inverse LBS canonicalization, it removes penetration layer by layer using a normal-direction projection step with a concrete thickness of 2 mm, fits a UDF implemented as a small MLP with positional encoding, and extracts the refined mesh with Marching Cubes using a threshold of 3 mm. On 4D-DRESS, it reports Intersection Ratio below $0.5\%$ in all categories, with $0.052\%$ for lower garments, $0.204\%$ for inner garments, and $0.485\%$ for outer garments (Vuran et al., 2 Aug 2025).

MulayCap addresses monocular video without pre-scanning by using a multi-layer geometry pipeline and a multi-layer texture pipeline. Geometry is decomposed into a body mesh layer plus garment piece layers, with garment reconstruction driven by Garment-from-Video and a cloth simulation model optimized by gradient descent. Texture is decomposed into shading and albedo layers, and the shading signal is used for shape-from-shading refinement while a fixed garment albedo atlas is fused over time. The method is explicitly motivated by the claim that layered modeling yields more robust and more editable performance capture than treating the person as one deforming surface (Su et al., 2020).

A different reconstruction objective appears in body-under-clothing recovery. ClothHMR treats clothing as an outer layer that should be trimmed away to expose a body-consistent silhouette. Its Clothing Tailoring module performs body semantic estimation, body edge prediction, and cloth cutting to produce a tailored body image, after which an FHVM-based mesh recovering module fits an SMPL mesh using joints, depth, and silhouette cues from a single foundational human visual model. The paper explicitly frames this as a layered viewpoint in which clothing is an outer perturbation of the body silhouette, but it does not reconstruct a separate garment layer or perform full layered cloth-body physical modeling (Gao et al., 19 Dec 2025).

5. Layer-wise generation, composition, and editing

Layer-wise generation systems treat clothed human synthesis as progressive dressing. HumanLiff generates a minimal-clothed human in canonical space and then progressively generates clothing layers using a unified diffusion process over tri-plane features. Its conditioning mechanism preserves 3D spatial structure by hierarchically fusing multi-scale 3D condition features from the previous layer, rather than compressing the condition into a 1D code. The method is evaluated on the layer-wise SynBody and TightCap datasets and reports substantial gains over EG3D, EVA3D, and Rodin in L-PSNR and FID, while tri-plane shift improves local detail without increasing total tri-plane parameters (Hu et al., 2023).

TELA also uses progressive inside-out generation, but with separate NeRFs for the minimal-clothed body and each garment layer. Its stratified compositional rendering uses transparency to divide each ray into layer-specific strata, so the current cloth NeRF is responsible only for points in its own region. A dual SDS loss supervises both the composited person-wearing-garment rendering and a cloth-only rendering, which is used to decouple the garment from the body. The method reports FID 125.9 compared with DreamWaltz at 142.6, and user-study preferences of 71.46% for overall clothed human quality, 77.92% for cloth quality, and 73.18% for consistency with text (Dong et al., 2024).

HumanCoser pursues physically layered generation through separate networks for the inner body and each clothing layer, a dual-representation decoupling framework, and a multi-layer fusion volume rendering rule in which only the most relevant layer at each sample contributes. An SMPL-driven implicit deformation network enables clothing reuse across identities, and the system reports FID 298.54 and CLIP score 31.61 while supporting clothing transfer and layered animation (Wang et al., 2024).

In feed-forward generation, LayerAvatar introduces a layered UV feature plane in a Gaussian-based representation. It uses a three-layer UV space with the innermost human body in Layer 1, hair and shoes in Layer 2, and top and bottom clothes in Layer 3. Separate body constraints are added because the innermost layer is severely occluded. On THuman2.0, the full layered pipeline reports FID 12.50 and KID 9.39, outperforming single-layer alternatives (Zhang et al., 8 Jan 2025).

6. Dynamics, inter-layer interaction, and animation

Once layers are made explicit, articulation and physics become central. UNIF represents a clothed human as a union of part-wise SDFs and learns part structure from motion rather than from SMPL part labels. Its Partition-from-Motion, bone-centered initialization, bone limit loss, section normal loss, adjacent part seaming, and “Competing Parts” blending are designed to preserve continuity and suppress extra surfaces in articulated clothed humans. The method is evaluated on CAPE and ClothSeq and is described as more suitable for loose clothes and articulated motion than SMPL-label-driven part models (Qian et al., 2022).

For garment-garment interaction specifically, ClothCombo addresses the draping of arbitrary combinations of clothes. It first computes a topology-agnostic cloth embedding with DiffusionNet, then predicts single-garment draping conditioned on body shape and pose, and finally applies a GNN-based untangling network that models pairwise inter-garment forces and predicts per-vertex displacements to resolve interpenetration. Multi-collision and distance losses enforce layer ordering, and the method is trained with randomized layer counts from 1 to 6 while showing generalization to 7 and 8 layers at test time (Lee et al., 2023).

Avatar systems have recently adopted explicit body-cloth separation in Gaussian splatting. PICA uses two distinct layers—an underlying body layer and an outer clothing layer—each with its own mesh template and its own 3D Gaussians. Body and clothing have separate deformation MLPs, a collision loss keeps clothing outside the body, and HOOD-based cloth simulation drives garment dynamics for novel poses. The method reports around 12 fps at test time and explicitly targets garment sliding, loose dresses, and virtual try-on, which are difficult for single-layer pose-conditioned models (Peng et al., 2024). Cloth-HUGS similarly factorizes performers into separate Gaussian layers for body and cloth in a shared canonical space, uses cloth-specific triplanes and physics-inspired constraints, and introduces a depth-aware multi-pass renderer; it reports real-time rendering at over 60 FPS and improvements in LPIPS and geometric fidelity over monolithic Gaussian avatar baselines (Mubashshira et al., 17 Apr 2026).

7. Limitations and unresolved issues

Despite the diversity of layered formulations, several difficulties recur. Hidden-layer inference remains substantially harder than visible-layer classification in layered segmentation, particularly when the system must infer occluded garment parts from visible geometry and body shape (Garavaso et al., 7 Aug 2025). Collision handling, while central, is usually geometric or weakly physics-inspired rather than a full physical cloth simulator, and several systems note that highly complex outfits, multiple overlapping garments, very thin or translucent materials, or cloth-cloth interactions remain challenging (Mubashshira et al., 17 Apr 2026).

Dependence on parametric body priors is another persistent constraint. ReMu relies on SMPL-X fits and nearest-vertex skinning weights for canonicalization (Vuran et al., 2 Aug 2025). HiLo and ARCH++ use parametric body models as internal scaffolds or correspondence priors rather than reconstructing explicit body-cloth layers (Yang et al., 2024). ClothHMR attributes failures on unconventional poses and extreme obesity to SMPL’s limited expressive power and proposes non-parameterized models as future work (Gao et al., 19 Dec 2025).

A final divide separates explicit and implicit layering. Methods such as PIFu, FastHuman, HiLo, ARCH++, and MetaAvatar reconstruct detailed clothed humans and often benefit from body-conditioned priors, canonicalization, or pose-aware deformation, but they still model the result primarily as a single clothed surface rather than as separate body and garment layers (Saito et al., 2019). This suggests that clothed human layering is not a single technique but a design axis: at one end are methods that infer only the outer clothed shell, and at the other are systems that explicitly represent overlapping semantics, ordered geometry, reusable garments, or independently animated body and clothing layers.