NI-Tex: Non-isometric Image-based Garment Texture Generation

TL;DR

NI-Tex addresses the challenge of limited texture diversity under non-isometric deformations by creating 3D Garment Videos and applying Nano Banana–driven topology edits to generate robust image–garment training pairs. A feedforward dual-branch network with Multi-Channel Aligned Attention synthesizes high-quality PBR textures that transfer across varied poses and topologies. An uncertainty-guided iterative baking pipeline combines multi-view predictions into seamless, production-ready textures, achieving industry-level texture fidelity. The approach demonstrates strong robustness and competitive quantitative performance, enabling practical deployment for 3D garment design and virtual try-on applications.

Abstract

Existing industrial 3D garment meshes already cover most real-world clothing geometries, yet their texture diversity remains limited. To acquire more realistic textures, generative methods are often used to extract Physically-based Rendering (PBR) textures and materials from large collections of wild images and project them back onto garment meshes. However, most image-conditioned texture generation approaches require strict topological consistency between the input image and the input 3D mesh, or rely on accurate mesh deformation to match to the image poses, which significantly constrains the texture generation quality and flexibility. To address the challenging problem of non-isometric image-based garment texture generation, we construct 3D Garment Videos, a physically simulated, garment-centric dataset that provides consistent geometry and material supervision across diverse deformations, enabling robust cross-pose texture learning. We further employ Nano Banana for high-quality non-isometric image editing, achieving reliable cross-topology texture generation between non-isometric image-geometry pairs. Finally, we propose an iterative baking method via uncertainty-guided view selection and reweighting that fuses multi-view predictions into seamless, production-ready PBR textures. Through extensive experiments, we demonstrate that our feedforward dual-branch architecture generates versatile and spatially aligned PBR materials suitable for industry-level 3D garment design.

Figures (15)

• Figure 1: NI-Tex takes an image prompt and a target garment geometry as input, and generates high-quality PBR textures that faithfully transfer the textures and patterns from the input image to the target garment. Unlike other existing methods, the generation quality of NI-Tex does not degrade for challenging image-garment pairs with strong topological and geometric differences, enabling superior flexibility.
• Figure 2: Texture generation becomes unreliable when the image prompt and the given mesh exhibit topology inconsistency in Hunyuan3D hunyuan3d2025hunyuan3d and Meshy 6 Preview.
• Figure 3: Overview of NI-Tex. In the data space (top), we construct our non-isometric training dataset from 3D Garment Videos by randomly selecting two frames, one as the condition 3D frame and the other as the supervision 3D frame, to enhance the model’s generalization across human poses, geometric deformations, and lighting variations. To further improve robustness to different garment topology, we apply Nano Banana for image editing on randomly rendered views of the condition 3D frame. In the pipeline (bottom), we render the input texture-less mesh to obtain normal and position maps as geometric constraints and employ a dual-branch architecture to achieve non-isometric PBR view generation. Finally, view selection and Uncertainty Quantification (UQ) are used to iteratively bake the results across multiple viewpoints.
• Figure 4: We use Nano Banana to edit garment topology while preserving texture consistency (first row). We ensure category-wise and inner–outer texture consistency to avoid texture swaps or layering confusion (second and third rows). Additional human body generations (fourth row) are acceptable, as image–garment training improves NI-Tex’s understanding of garment textures.
• Figure 5: We render each garment mesh from six viewpoints and bake the results to check coverage. Self-occlusion under orthogonal views leaves many regions missing. Coverage-based view selection improves coverage but still misses small areas, while ours achieves full mesh coverage.