Transcending Parametric Templates for Clothed Human Reconstruction

This direction is merged into another direction: non-rigid geometry modeling. Its last version stays readable and citable here.

Version 1 of 157 papers61 conceptsfrozen 2026-10-04

For years, the cornerstone of 3D human digitizing has been the parametric body model—mathematical templates like SMPL that define a "standard" human shape controlled by a few dozen skeletal and shape parameters. While these models provide a robust structural anchor, they suffer from a fundamental "parametric bottleneck": they are designed to represent naked or tight-clothed bodies.

When a person wears a flowing skirt, a bulky coat, or has long hair, the rigid topology of these templates fails. Recent research has shifted toward hybrid and implicit representations that use parametric models as a guide rather than a cage, allowing for the reconstruction of complex, non-rigid human geometry.

The Parametric Bottleneck

The primary limitation of traditional templates is their fixed topology. As noted by Mana Masuda and colleagues in their work on Generalizable Neural Human Renderer, these models cannot capture geometry that deviates significantly from the base mesh. Because the mesh assumes a tight fit to the human form, it essentially "erases" the volume of loose clothing or accessories.

This constraint creates a performance gap. Leonid Pishchulin and Bernt Schiele observed that detectors trained on these "tight-fitting" synthetic models struggle when encountering diverse real-world apparel like coats or skirts Learning people detection models from few training samples. To bridge this gap, the field has moved toward representations that can "breathe" beyond the skeleton.

From Rigid Meshes to Neural Fields

To capture the "non-conforming" geometry of real life, researchers have turned to Neural Radiance Fields (NeRFs) and volumetric representations. Instead of fitting a mesh, these methods represent the human as a continuous volume.

Gül Varol and Ivan Laptev pioneered this with BodyNet, which uses a voxel grid to infer shape directly from images, bypassing the need for parametric regression which often loses subject-specific details. More recently, Wei Jiang 0034 and Anurag Ranjan introduced NeuMan, which uses two separate NeRFs—one for the person and one for the scene—to reconstruct photorealistic humans in new poses from casual video, capturing the fine-grained interactions that rigid models miss.

Modeling Pose-Dependent Dynamics

A major challenge in moving beyond rigid templates is that clothing doesn't just have a static shape; it deforms dynamically based on pose. A sleeve wrinkles when an arm bends; a skirt swings when a person turns.

Chunjin Song and Helge Rhodin addressed this with Pose Modulated Avatars. By adaptively adjusting frequency encodings based on skeletal pose, their framework captures dynamic wrinkles while maintaining smooth skin surfaces. Similarly, Wenfeng Song and Aimin Hao developed CtrlAvatar, which uses "invertible delta networks" to decouple body geometry from texture, allowing for real-time rendering of non-rigid deformations like facial expressions and clothing folds.

Generalization and the End of Optimization

Early neural rendering methods required hours of optimization for every new subject. The current frontier is generalizable priors—models that can reconstruct a clothed human from a single image or sparse views instantly.

Methods like HFGaussian by Arnab Dey use 3D Gaussian Splatting to achieve real-time reconstruction (25 FPS) of unseen subjects. Meanwhile, Cheng Peng and Yebin Liu introduced the Parametric Gaussian Human Model, which pre-trains on massive datasets to bridge the gap between slow, high-quality multi-view systems and fast but blurry monocular ones.

Bypassing the Mesh with Diffusion

The most recent shift involves using generative diffusion models to bypass 3D meshes entirely. Instead of trying to "fit" a person into a 3D space, these models "hallucinate" consistent 3D views.

Ethan Weber and Yash Kant demonstrated this with Pippo, a multi-view diffusion transformer that generates 1K-resolution turnaround images from a single photo. By using minimal spatial anchors instead of a full SMPL mesh, Pippo maintains geometric consistency while capturing the full complexity of hair and diverse fashion styles that have long eluded parametric templates.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_transcending_parametric_templates_for_cl_4d7ba252,
  title = {Transcending Parametric Templates for Clothed Human Reconstruction},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/transcending-parametric-templates-for-clothed-human-reconstruction/rdr_4d7ba25260654d5984a5fd7f001f2fc3},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_transcending_parametric_templates_for_cl_4d7ba252_v1,
  title = {Transcending Parametric Templates for Clothed Human Reconstruction},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/transcending-parametric-templates-for-clothed-human-reconstruction/rdr_4d7ba25260654d5984a5fd7f001f2fc3/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}