Garment Geometry Image (GGI)

• Garment Geometry Image (GGI) is a UV-based representation that encodes dense 3D garment geometry and semantic details for efficient, topology-aware mesh reconstruction.
• It integrates per-panel mesh data, stitching logic, and normalized UV mapping through rasterization and interpolation to bridge 2D pattern reasoning with 3D structure.
• GGI pipelines offer real-time synthesis and high seam precision while being mesh-resolution independent, though they may smooth out high-frequency wrinkle details.

The Garment Geometry Image (GGI) is a multidimensional, UV-based array representation encoding the geometric and semantic structure of 3D garment surfaces. GGIs unify 2D sewing pattern reasoning with dense 3D geometry by mapping per-panel mesh data, semantic information, and stitching logic into regularly gridded images suited to neural network prediction and efficient mesh reconstruction. GGIs serve as an intermediary for garment generation, editing, and simulation, facilitating direct, topology-aware synthesis and inverse processes that assemble watertight garment meshes from compact visual representations (Pham et al., 19 Mar 2026, Dumoulin et al., 4 Apr 2025, Li et al., 2 Apr 2025).

1. Formal Definition and Structure

A GGI encodes the surface geometry of each garment panel through dense rasterization in UV space, often complemented with semantic and stitching annotations. Across recent frameworks:

• GarmageNet defines the per-panel GGI $G_i \in \mathbb{R}^{H \times W \times C}$ as

$$G_i[x,y] = (x_i[x,y],\,y_i[x,y],\,z_i[x,y],\,\alpha_i[x,y]),$$

where $(x_i, y_i, z_i)$ are normalized, centered 3D coordinates within the panel’s bounding box, and $\alpha_i[x,y] \in \{0,1\}$ marks the pixel as inside or outside the pattern domain $P_i$ (Li et al., 2 Apr 2025).

• SwiftTailor generalizes to a global UV map partitioned into disjoint panel regions $U_i$, with images:
  • $G_{\text{sem}} \in [0,1]^{H \times W \times 3}$ (panel type semantics)
  • $G_{\text{geo}} \in [0,1]^{H \times W \times 3}$ (normalized 3D geometry)
  • $G_{\text{st}} \in [0,1]^{H \times W \times K}$ (stitching tags)
  • —on a shared regular grid (Pham et al., 19 Mar 2026).
• D-Garment operates on a garment template mesh with precomputed low-distortion UV mapping $\phi:\mathcal{T} \rightarrow [0,1]^2$. The GGI records, per pixel, the 3D vertex displacement from the template: $D[i,j] = M^{\phi(i,j)} - T^{\phi(i,j)}$ (Dumoulin et al., 4 Apr 2025).

Most implementations discretize UV space into $H \times W$ (typically $256^2$ to $512^2$) grids, with three geometric channels (XYZ or offsets) and, optionally, one or more semantic or mask channels.

2. Forward Mapping: 3D Mesh to Geometry Image

The forward GGI construction interpolates mesh attributes onto the UV grid. For each panel:

• Rasterize or pack mesh faces into the UV domain $[0,1]^2$ or per-panel patch $P_i$.
• For each pixel $(i,j)$:
  • Map to UV via $u_x = x/(H-1)$, $v_y = y/(W-1)$.
  • If within domain ($\chi_{P_i}(u_x, v_y) = 1$), barycentrically interpolate the 3D position and, if required, surface normal.
  • Center and normalize coordinates relative to the panel’s geometric bounding box.

In SwiftTailor (Pham et al., 19 Mar 2026), normalization to $[0,1]$ is performed:

$$x' = \frac{x - x_{\min}}{x_{\max} - x_{\min}}, \quad y', z' \text{ analogously}$$

For non-vertex UV positions, barycentric interpolation is used, ensuring coherent geometry and non-overlapping panel rasterization.

D-Garment registers the per-pixel 3D offset from the mesh template, i.e., $D[i,j] = M^{\phi(i,j)} - T^{\phi(i,j)}$, via bilinear splatting in UV space (Dumoulin et al., 4 Apr 2025).

3. Inverse Mapping: Mesh Reconstruction and Stitching

Mesh recovery from GGIs proceeds in two main stages:

• Remeshing: Interpret $G_{\text{geo}}$ (or per-panel $G_i$/$D$) as a height field over the pixel grid. For each occupied $2 \times 2$ cell, create triangular faces according to occupancy, with diagonal selection based on minimal 3D length.
• Dynamic Stitching: Boundary correspondence is established via $G_{\text{st}}$ or neural matching (e.g., PanelJigsaw in GarmageNet). Stitch-pair curves or point sets are aligned (often via Dynamic Time Warping), and per-panel boundary vertices are unified using Disjoint Set Union. Dual merging resolves overlapping vertices, yielding watertight mesh topology.

In GarmageNet, additional refinement fits matched pairs as piecewise-linear curves for precise sewing edges (Li et al., 2 Apr 2025). This technique admits direct pipeline absorption into industrial cloth simulation systems.

4. Learning and Synthesis with GGIs

Neural generative pipelines leverage GGIs as structured scene representations:

• SwiftTailor: The GarmentSewer Dense Prediction Transformer (DPT) predicts $G_{\text{geo}}$ and $G_{\text{st}}$ from a semantic image, optimized with edge-aware losses and Chamfer distance. PatternMaker predicts the input sewing pattern tokens and panel layouts (Pham et al., 19 Mar 2026).
• D-Garment: A VAE encodes GGIs into low-dimensional latents; a U-Net diffusion model synthesizes garment deformations in the latent space, conditioned on body shape, motion, and cloth material via cross-attention. The process is mesh-resolution independent (Dumoulin et al., 4 Apr 2025).
• GarmageNet: A dual-stage diffusion pipeline first denoises global panel topology, then synthesizes latent codes for per-panel GGIs using Transformers, with CLIP or pattern-conditioned inputs. GGIs are decoded at inference and stitched by learned boundary matchers (Li et al., 2 Apr 2025).

Pseudocode Example: SwiftTailor GGI and Mesh Reconstruction

pattern = PatternMaker(image_or_text_input) repacked_layout = LayoutPacking(pattern) G_sem = SemanticImage(pattern, repacked_layout) G_st = StitchingImage(pattern, repacked_layout) G_geo = GeometryImage(pattern, repacked_layout) G_geo_hat = GarmentSewer(G_sem) panel_meshes = Remeshing(G_geo_hat) # Alg. 5 in paper final_mesh = StitchPanels(G_st, pattern, panel_meshes) # Alg. 6 in paper

(Pham et al., 19 Mar 2026)

5. Semantic and Sewing Relationship Modeling

Stitching logic is encoded either as dedicated image channels (GGI stitching map) or via learned pairwise matchers (PanelJigsaw):

• GGI sewing maps: $G_{\text{st}}$ colors boundary pixels to identify stitch-pairs; matching stitches share color codes. This enables explicit, easily visualized seam associations (Pham et al., 19 Mar 2026).
• Learned correspondence (GarmageNet): PanelJigsaw extracts panel boundary points, encodes geometric and UV features via PointNet++/PointTransformer, classifies and matches points via Sinkhorn/Hungarian algorithms to produce a permutation matrix $X\in\{0,1\}^{M\times M}$, then fits stitch curves (Li et al., 2 Apr 2025).

This explicit correspondence is principal to automation of mesh assembly, simulation readiness, and progressive garment editing.

6. Quantitative Evaluation and Practical Considerations

GGI-based pipelines support comprehensive, fine-grained evaluation:

Metric	Value (GarmageNet)	Value (SwiftTailor/D-Garment)
Per-pixel GGI MSE	$1.2 \times 10^{-4}$ (Li et al., 2 Apr 2025)	—
3D Chamfer distance (recon error)	$\sim0.35$ mm (Li et al., 2 Apr 2025)	D-Garment beats SOTA in $CD$ (Dumoulin et al., 4 Apr 2025)
Simulation success (automated, no edits)	$58.5$ % (Li et al., 2 Apr 2025)	—
Inference time (GGI pipeline)	$\sim0.02$ s (Pham et al., 19 Mar 2026)	7.5 FPS synthesis (Dumoulin et al., 4 Apr 2025)
Baseline (2D pattern → sim mesh)	$30-50$ s (Pham et al., 19 Mar 2026)	—

GGI architectures attain real-time or near-real-time synthesis rates, outperforming serialized 2D pattern + simulation pipelines by over $10\times$ (Pham et al., 19 Mar 2026). Per-panel and global geometric losses, as well as semantic and stitching supervisions, yield sub-millimeter geometric fidelity and high seam precision.

7. Benefits and Limitations

Benefits:

• Efficiency: Drastic reduction in mesh synthesis and garment assembly times compared to physics-based simulation or serialized pattern modeling (Pham et al., 19 Mar 2026).
• Directness: Explicit 3D geometry and sewing semantics embedded, facilitating visualization, editability, and transparent downstream workflows (Li et al., 2 Apr 2025).
• Mesh-resolution independence: UV-based GGI mapping enables transfer across arbitrary mesh tessellations without retraining (D-Garment) (Dumoulin et al., 4 Apr 2025).
• Generalization: Modular pattern-to-GGI pipelines support multimodal conditioning (text, image, sketch) (Li et al., 2 Apr 2025).

Limitations:

• Wrinkle Resolution: Synthesis pipelines tend to produce smoothed geometry; high-frequency wrinkling often requires post-refinement (Pham et al., 19 Mar 2026).
• Dataset Dependence: Poor generalization to out-of-distribution or highly occluded silhouettes not represented in training (Pham et al., 19 Mar 2026).
• Resolution trade-off: UV grid size directly affects memory cost vs. geometric detail (Pham et al., 19 Mar 2026).

A plausible implication is that GGI representations may catalyze further research in high-fidelity, interpretable, and simulation-ready digital garment workflows, with active cross-pollination from related domains such as generative modeling, computational geometry, and industrial apparel engineering.