uv mapping

Existing 3D generation methods suffer from structural-textural misalignment and limited fidelity due to stage-wise, heterogeneous modeling of geometry and texture. To address this, we propose a unified generative framework based on a bidirectionally invertible 2.5D latent space: multi-view RGB, normal, and coordinate maps are jointly embedded into a shared latent representation, enabling end-to-end, text- or image-conditioned 3D synthesis. We introduce the first 2.5D latent variable formulation, uniquely supporting joint optimization of geometry and appearance. Furthermore, we design a lightweight 2.5D-to-3D refinement decoder that significantly improves texture consistency under geometric guidance. Our method achieves state-of-the-art performance across text-to-3D and image-to-3D benchmarks, yielding a +4.2 dB PSNR gain in geometry-guided texture reconstruction and markedly enhanced structural-color consistency.

Existing methods struggle to simultaneously achieve high-fidelity facial geometry reconstruction and consistent fixed-topology meshes in unconstrained, in-the-wild scenarios. To address this challenge, this work proposes a feedforward multi-view facial reconstruction framework that introduces, for the first time, a mask-aware UV-space neural fusion mechanism. This approach replaces heuristic topological optimization by directly generating high-quality, fixed-topology face meshes in the standard UV space. The method leverages VGGT and Pixel3DMM to extract multi-view point maps and UV correspondences, and enhances generalization through geometry-to-geometry cross-view fusion. Experiments demonstrate that the proposed method achieves state-of-the-art accuracy on multiple public benchmarks and real-world in-the-wild datasets, completing full reconstruction from 16 input views in under three seconds on a single RTX 4090 GPU.

Current 3D texture generation heavily relies on manual UV mapping—time-consuming and lacking semantic awareness and visibility considerations. To address this, we propose the first unsupervised, differentiable UV parameterization framework that jointly incorporates semantic consistency and visibility awareness. Our method (1) achieves semantically coherent UV chart decomposition via mesh semantic segmentation and cross-shape semantic alignment; (2) introduces ambient occlusion (AO)-weighted soft seam optimization to implicitly guide cuts toward low-visibility regions; and (3) designs an end-to-end trainable backbone that jointly optimizes UV parameterization and seam distribution. Quantitative and qualitative evaluations across multiple benchmarks demonstrate that our approach significantly reduces visible seam artifacts and substantially improves downstream texture generation quality and visual naturalness. This work establishes a new paradigm for automated, high-fidelity 3D content generation.

Existing UV mapping methods operate pixel-wise, lacking global anatomical consistency and thus producing implausible or locally inaccurate mappings. This work proposes Pose-Constrained Continuous Surface Embeddings (PC-CSE), the first approach to explicitly incorporate real-time estimated full-body pose as a strong geometric constraint into UV mapping. PC-CSE achieves this via continuous surface embedding, pose-guided pixel-to-vertex correspondence, and multi-scale feature alignment—jointly ensuring global coherence and local fidelity. Trained and evaluated end-to-end on DensePose COCO, PC-CSE significantly reduces invalid mapping rates and improves anatomical plausibility and structural consistency of UV maps. Experiments demonstrate that full-body pose constraints outperform upper-body-only constraints and that PC-CSE is compatible with arbitrary 2D pose estimators. Furthermore, our analysis uncovers systematic annotation inconsistencies in the DensePose dataset—highlighting a previously underreported limitation in current benchmarks.

Existing methods rely on UV mapping for multi-view image-to-3D-texture reprojection and inpainting, suffering from topological ambiguity that induces geometric inconsistency and texture distortion. This paper proposes a two-stage 3D texture generation framework that bypasses explicit UV parameterization and directly models texture within a unified 3D functional space. Key contributions include: (1) the first continuous voxel-based representation of Texture Functions (TFs); (2) the construction of the first Large Texture Model (LTM), a scalable, text-conditioned foundation model for 3D texture generation; and (3) a LoRA-adapted Diffusion Transformer (DiT) architecture enabling joint multi-view optimization. Experiments demonstrate state-of-the-art performance in visual fidelity, geometric consistency, and cross-shape generalization. The framework enables fully automatic, high-quality, and scalable 3D texture synthesis without manual UV unwrapping or mesh editing.

Monocular 3D clothed human reconstruction is often hindered by scarce texture data, inaccurate geometric priors, and supervision bias inherent to single-modality learning, leading to suboptimal reconstruction quality. To address these limitations, this work proposes a geometry–texture collaborative reconstruction framework. We construct a large-scale dataset comprising over 15,000 textured 3D human scans and introduce a multi-source texture synthesis strategy, a region-aware shape extraction module, and a Fourier-based geometric encoding mechanism. A dual-branch U-Net architecture is further designed to effectively fuse geometry and texture features. By transcending the constraints of single-modality supervision, our method achieves state-of-the-art performance across multiple benchmarks and in-the-wild images, enabling high-fidelity, high-quality 3D reconstruction of clothed humans from a single image.

This work addresses the challenges of view inconsistency and missing textures in occluded regions inherent in multi-view texture generation, as well as the limited generalization and inability of conventional UV inpainting methods to leverage 2D diffusion priors. To overcome these limitations, we propose a unified framework that integrates multi-view image generation priors with a UV-space generative model. Our approach simultaneously inpaints occluded regions and enforces multi-view consistency directly in UV space, effectively combining the rich semantic priors of 2D diffusion models with the geometric coherence of UV representations. Experimental results demonstrate that our method significantly improves texture quality in both unseen and view-conflicting regions, outperforming existing approaches and achieving, for the first time, effective synergy between multi-view generative priors and UV-space generative modeling.

Existing single-image multi-view face generation methods suffer from cross-view geometric inconsistency due to the absence of explicit 3D structural constraints. This work proposes a dual-stream diffusion framework that jointly synthesizes multi-view RGB images and 3D facial geometry by introducing a viewpoint-invariant UV position map as a shared geometric representation. To enforce mutual consistency between appearance and geometry, the method incorporates a geometry-guided attention alignment loss leveraging a shared attention mechanism. Evaluated on the RenderMe-360 and NeRSemble datasets, the proposed approach significantly outperforms existing methods in both visual fidelity and cross-view geometric consistency, while also enabling more efficient 3D reconstruction.

This work addresses the challenging problem of reconstructing high-fidelity, renderable 3D facial models from only a few uncalibrated face images. We propose the first co-optimization framework integrating Gaussian splatting with explicit triangular meshes. Methodologically, we introduce semantic-segmentation-guided geometric alignment and soft mesh constraints to ensure accurate neutral-pose modeling; design a view-dependent mapping from Gaussian points to texture space for generating 4K neural textures; and achieve illumination-decoupled albedo extraction with cross-illumination robust training. Our contributions are threefold: (1) high-quality meshes and textures are generated from Gaussian representations without modifying standard graphics pipelines; (2) fine-grained, animation- and relighting-ready facial assets are produced from merely 11 input images; and (3) strong generalization and practical utility are demonstrated in text-driven 3D face generation tasks.

Existing 3D quality assessment metrics (e.g., Chamfer Distance) exhibit poor correlation with human perception, while rendering-based learning methods suffer from view bias, incomplete structural coverage, and inadequate modeling of authentic distortions. To address these limitations, this work proposes the first end-to-end fidelity assessment method specifically designed for textured 3D meshes. We introduce a 3D latent geometric network that jointly encodes geometric structure and surface color features—bypassing rendering-induced artifacts. Furthermore, we construct the first high-quality, human-annotated dataset of textured 3D models exhibiting realistic distortions. Extensive experiments demonstrate that our method significantly outperforms state-of-the-art geometry- and rendering-based baselines in perceptual consistency, robustness, and generalization across diverse distortion types. On multiple benchmarks, it achieves up to a 12.6% improvement in Spearman rank-order correlation coefficient (SROCC).