Garment Geometry Image

Updated 3 July 2026

Garment Geometry Images are 2D arrays that encode 3D garment surface properties using UV mapping, capturing geometry, normals, and semantic details.
They enable efficient conversion of 2D design inputs to detailed 3D models, supporting synthesis, re-texturing, and controllable garment editing.
Modern implementations use neural architectures for accurate, simulation-ready mesh assembly with improved topology and stitching consistency.

A garment geometry image is a structured, image-based encoding of garment surface geometry—often in UV or panel-aligned coordinates—used to synthesize, reconstruct, analyze, or edit 3D clothing from photographic, sketch, parametric, or multimodal input. Serving as an intermediary between 2D design representations and high-fidelity 3D garment models, garment geometry images enable both efficient neural processing and robust recovery of semantically structured, physically plausible, stitch-compatible garment meshes. Modern approaches leverage geometry images for tasks ranging from garment generation and re-texturing, to simulation-ready mesh assembly, and controllable garment editing.

1. Core Definitions and Formalism

A garment geometry image (GGI) is a multi-channel, spatially regular 2D array encoding geometric, semantic, and topological properties of a garment surface. Its channels may include 3D surface coordinates, surface normals, semantic panel type annotations, stitching relationships, or occupancy/mask values. The parameterization is typically aligned to panel UV coordinates $[0,1]^2$ or a canonical, packed UV atlas for full garments. For point $(u,v)$ in UV space, the geometry image may store:

$G_\mathrm{geo}(u,v)$ : 3D surface position $(x,y,z)$
$G_\mathrm{norm}(u,v)$ : surface normal $\mathbf{n}$
$G_\mathrm{sem}(u,v)$ : integer/one-hot encoding of panel type
$G_\mathrm{stitch}(u,v)$ : seam/stitch identifier

GGIs are constructed by rasterizing a 3D mesh (or implicit representation) onto a dense UV grid, with interpolated values at each valid $(u,v)$ . Inverse mapping uses GGI samples to reconstruct a mesh via explicit remeshing and dynamic stitching of UV-adjacent panel boundaries (Pham et al., 19 Mar 2026, Li et al., 2 Apr 2025).

2. Geometry Image Construction and Inversion

The construction of a GGI from a garment mesh or implicit surface uses mesh UV coordinates to sample 3D vertex positions and auxiliary properties onto a grid. For a panel mesh $M_i=(V_i,T_i)$ with UV parameterization over patch $(u,v)$ 0:

For each grid cell $(u,v)$ 1 at $(u,v)$ 2, the GGI channel values are set as

$(u,v)$ 3

where $(u,v)$ 4 and $(u,v)$ 5 are normalization parameters, and $(u,v)$ 6 is the binary mask for panel $(u,v)$ 7 (Li et al., 2 Apr 2025).

Inversion to mesh proceeds by identifying "active" grid cells ( $(u,v)$ 8), reconstructing 3D vertex positions by de-normalizing $(u,v)$ 9, and connecting adjacent grid cells via triangulation to build per-panel meshes. Sewn garment surfaces are formed by matching panel edge pixels with identical seam identifiers and merging/averaging corresponding 3D vertices (Pham et al., 19 Mar 2026, Li et al., 2 Apr 2025).

3. Network Architectures for Geometry Image Prediction

Neural architectures for predicting GGIs from sketches, photos, sewing patterns, or parametric templates typically use encoder-decoder or transformer-based networks to densify geometry in UV/image space:

GarmentSewer (Pham et al., 19 Mar 2026): a vision transformer (ViT-L) with a multi-scale convolutional decoder takes as input semantic and stitch images to predict $G_\mathrm{geo}(u,v)$ 0.
GarmageNet (Li et al., 2 Apr 2025): a latent diffusion transformer synthesizes latent panel-wise geometry images $G_\mathrm{geo}(u,v)$ 1 which are decoded by a VAE.
UV-guided occupancy networks (Luo et al., 2024): MLPs combine pixel-aligned and geometry-aware features to regress surface occupancy in 3D or boundary distances in UV space.
HiGarment (Guo et al., 29 May 2025): encodes flat sketches into visual features and fuses them with text and fabric retrieval embeddings to control cross-modal garment geometry image synthesis.

Losses for GGI prediction include edge-aware L1 (for boundary accuracy), structure-aligned Chamfer (for seam consistency), normal regularization, and diffusion denoising objectives in latent or pixel space.

4. Applications and Task-Specific Adaptations

Garment geometry images are foundational in a spectrum of garment-related tasks:

3D Garment Generation: SwiftTailor uses GGIs as a dense assembly for fast, accurate 3D mesh reconstruction and dynamic stitching, outperforming physics-based simulation in speed and accuracy (Pham et al., 19 Mar 2026).
Pattern-to-Geometry Synthesis: GarmageNet generates panel GGIs from raw sewing patterns or multimodal inputs, enabling automatic panel assembly and downstream simulation (Li et al., 2 Apr 2025).
UV Map and Normal-Based Editing: Normal-guided networks learn geometry-aware UV-to-image mappings for isometric re-texturing and surface manipulation directly in the GGI domain (Jafarian et al., 2023).
Sim-to-Real and Topology-Aware Reconstruction: ReWeaver predicts panel geometry images and seam adjacency from few views, ensuring watertight surfaces and simulation-compatibility (Li et al., 23 Jan 2026).
Cross-Modal and Text-Guided Synthesis: Flat-sketch-to-image frameworks leverage geometry images as structure guides while harmonizing with textual/material cues for photorealistic sample generation (Guo et al., 29 May 2025).

5. Quantitative Performance and Comparative Analysis

Experimental benchmarks consistently indicate that GGI-centric frameworks achieve superior mesh fidelity, topology correctness, and inference speed relative to explicit mesh regression and unstructured 3D point methods:

SwiftTailor reduces Chamfer-based L1 error (MMD) by ∼23% and increases coverage by +27% versus competing sewing-based approaches, with inference times as low as 14.8s (Pham et al., 19 Mar 2026).
GarmentX, while operating in a compressed parametric garment space, leverages GGI-style pattern decoding to achieve 0% simulation failure and notably lower point-to-surface and Chamfer distance compared to mesh regression baselines (Guo et al., 29 Apr 2025).
GarVerseLOD demonstrates the value of hierarchical, LOD-driven GGI representation by facilitating state-of-the-art 3D reconstruction on in-the-wild garment images (Luo et al., 2024).
ReWeaver's topology accuracy and panel boundary IoU are nearly double those of previous multi-view garment methods, credited to GGI-based UV/curve mapping (Li et al., 23 Jan 2026).

6. Extensions and Future Directions

Recent trends extend the GGI concept to multi-layer garments, dynamic sequences, and emerging applications:

Multilayer and Temporal Modeling: LGN encodes layered garments as implicit fields with GGI-aligned coverage constraints, ensuring non-intersecting layers (Aggarwal et al., 2022); spatio-temporal GGI priors in diffusion-based mapping yield realistic, temporally coherent clothing animation (You et al., 27 Feb 2026).
Interactive and Production-Ready Pipelines: The editable parametric dimensions in GarmentX and the point-to-point seam prediction in GarmageNet enable user-in-the-loop and fully automated production workflows (Guo et al., 29 Apr 2025, Li et al., 2 Apr 2025).
Evaluation on Real Scans and Simulation: GGI-based methods are increasingly benchmarked on sim-to-real transferability, e.g., physical simulation precision and minimum penetration depth on body scans.

A plausible implication is that geometry images will remain central in garment digitization pipelines, bridging the gap between deep learning–friendly 2D data, parametric pattern representations, and the requirements of physically plausible, simulation-ready 3D garment models.

References:

(Guo et al., 29 May 2025, Pham et al., 19 Mar 2026, Li et al., 2 Apr 2025, Li et al., 23 Jan 2026, Aggarwal et al., 2022, Jafarian et al., 2023, You et al., 27 Feb 2026, Luo et al., 2024, Guo et al., 29 Apr 2025)