Mesh-Aware Gaussian UV Mapping

• Mesh-aware Gaussian UV mapping is a method that converts unstructured 3D Gaussian splats into a structured 2D UV representation, enhancing compatibility with standard graphics pipelines.
• It leverages spherical projection and optimal transport-inspired capacity allocation to improve UV slot utilization and boost rendering metrics such as PSNR and SSIM.
• Mesh awareness integrates geometric and semantic constraints, ensuring stable surface attachment, seamless deformation handling, and editable textures on dynamic models.

Searching arXiv for the cited papers to ground the article and confirm metadata. arxiv_search(query="(Kim, 21 Apr 2026) OR (Rai et al., 26 Feb 2026) OR (Srinivasan et al., 2023) OR (Jiang et al., 2024) OR (Rai et al., 3 Feb 2025)", max_results=10, sort_by="relevance") arxiv_search(query="Mesh-Aware Gaussian UV Mapping", max_results=10, sort_by="relevance") Mesh-aware Gaussian UV mapping denotes a family of representations and algorithms that place Gaussian primitives, Gaussian attributes, or Gaussian-derived surface signals into a UV domain so that unstructured splats can be organized, edited, compressed, animated, or rendered through a structured parameterization. Across recent work, the term covers several distinct regimes: spherical UV parameterization of unstructured 3D Gaussians into image-like tensors; direct optimization of Gaussian attributes in UV space for volumetric video; mesh-guided learning of Gaussian textures on deforming surfaces; and explicit embedding of Gaussian local frames or splats into the UV atlas of a template mesh for editing, relighting, and map extraction (Rai et al., 3 Feb 2025, Rai et al., 26 Feb 2026, Jiang et al., 2024, Baert et al., 9 Dec 2025). A central unifying idea is that UV organization converts permutation-invariant Gaussian sets into structured domains that admit 2D networks, atlas packing, codec compatibility, or conventional graphics operations, while mesh awareness determines how strongly the UV map respects topology, deformation, semantics, visibility, or surface attachment.

1. Conceptual foundations

In the UVGS formulation, an unstructured set of 3D Gaussians, $\mathcal{G}$, is mapped to a regular UV tensor $$\Phi(\mathcal{G}) \in \mathbb{R}^{H \times W \times K \times C}$$, where $H \times W$ is the UV grid, $K$ is per-slot Gaussian capacity, and $C$ is the number of feature channels (Kim, 21 Apr 2026). The original UVGS paper frames this as a way to transform 3DGS into a structured 2D representation using spherical mapping, so that Gaussian attributes such as position, scale, color, opacity, and rotation become channels in a UV image (Rai et al., 3 Feb 2025). This regularization is technically significant because vanilla 3DGS is discrete, unstructured, and permutation-invariant, whereas UV organization makes the representation compatible with ConvNets, autoencoders, VAEs, latent diffusion models, and standard image operations (Rai et al., 3 Feb 2025).

Mesh awareness, however, is not a single property. In some systems it means that Gaussians are physically attached to a surface and move with a deforming mesh; in others it denotes that UV coordinates are continuous, semantically aligned, visibility-aware, or locally coherent with surface neighborhoods. UV Gaussians uses a refined mesh as the 3D anchor for Gaussian textures predicted in UV space, so that Gaussians remain attached to the body surface and follow pose-dependent deformations (Jiang et al., 2024). GTAvatar embeds each canonical Gaussian primitive’s local frame into a patch in the UV space of a template mesh, enabling continuous editable material textures on a conventional UV domain (Baert et al., 9 Dec 2025). DeMapGS goes further by anchoring each Gaussian splat to a specific triangle face on a template mesh through barycentric coordinates and normal displacement, which makes the coupling to topology explicit (Zhou et al., 11 Dec 2025). By contrast, OT-UVGS explicitly notes that it does not use mesh structure or topology in the mapping step, even though its capacity-aware assignment is relevant to mesh-guided settings (Kim, 21 Apr 2026).

A concise distinction therefore emerges. UV-structured Gaussian mapping addresses storage, regularity, and compatibility with 2D processing; mesh-aware Gaussian UV mapping adds geometric or semantic constraints that make the UV domain correspond to a surface in a stable and editable way. This suggests that “mesh-aware” is best treated as a spectrum rather than a binary label.

2. From spherical UVGS to capacity-aware allocation

The baseline UVGS mapping is spherical projection: each Gaussian center is converted to spherical coordinates and projected independently to a UV grid (Rai et al., 3 Feb 2025). For a Gaussian center $(x_i,y_i,z_i)$, spherical conversion is described as

$$\rho_i = \sqrt{x_i^2 + y_i^2 + z_i^2},\quad \theta_i = \tan^{-1}(y_i, x_i),\quad \phi_i = \cos^{-1}\left(\frac{z_i}{\rho_i}\right),$$

followed by discretization onto a grid (Rai et al., 3 Feb 2025, Rai et al., 26 Feb 2026). UVGS stores the attributes

$$[x, y, z, r1, r2, r3, r4, s1, s2, s3, o, c1, c2, c3]$$

as a 14-channel map and resolves many-to-one assignments by retaining the Gaussian with highest opacity, with a multi-layer extension that stores the top-$K$ opacity values per pixel (Rai et al., 3 Feb 2025).

OT-UVGS reinterprets this mapping as a capacity-allocation problem under a fixed UV budget (Kim, 21 Apr 2026). The paper argues that deterministic spherical projection ignores the global Gaussian distribution, leaving many UV slots empty while causing frequent collisions in dense regions. Its optimal-transport-inspired solution computes angular coordinates, ranks all Gaussians globally in $\theta$ and $$\Phi(\mathcal{G}) \in \mathbb{R}^{H \times W \times K \times C}$$0, and maps the normalized ranks to UV:

$$\Phi(\mathcal{G}) \in \mathbb{R}^{H \times W \times K \times C}$$1

with slot indices

$$\Phi(\mathcal{G}) \in \mathbb{R}^{H \times W \times K \times C}$$2

If more than $$\Phi(\mathcal{G}) \in \mathbb{R}^{H \times W \times K \times C}$$3 Gaussians map to one slot, only the $$\Phi(\mathcal{G}) \in \mathbb{R}^{H \times W \times K \times C}$$4 with highest opacity $$\Phi(\mathcal{G}) \in \mathbb{R}^{H \times W \times K \times C}$$5 are retained (Kim, 21 Apr 2026).

The methodological contrast is precise. Spherical projection is local and pointwise; OT-UVGS is globally coupled via ranks; histogram equalization balances each marginal but lacks full coupling between $$\Phi(\mathcal{G}) \in \mathbb{R}^{H \times W \times K \times C}$$6 (Kim, 21 Apr 2026). The dominant computation in OT-UVGS is sorting in each angular dimension, giving overall complexity $$\Phi(\mathcal{G}) \in \mathbb{R}^{H \times W \times K \times C}$$7, and the method is presented as a drop-in replacement for spherical UVGS (Kim, 21 Apr 2026).

Empirically, the capacity-allocation view changes both utilization and rendering quality. On 184 object-centric scenes, UVGS reports PSNR $$\Phi(\mathcal{G}) \in \mathbb{R}^{H \times W \times K \times C}$$8, SSIM $$\Phi(\mathcal{G}) \in \mathbb{R}^{H \times W \times K \times C}$$9, and LPIPS $H \times W$0, while OT-UVGS reports PSNR $H \times W$1, SSIM $H \times W$2, and LPIPS $H \times W$3; on the Mip-NeRF full scene, UVGS reports PSNR $H \times W$4, SSIM $H \times W$5, and LPIPS $H \times W$6, while OT-UVGS reports PSNR $H \times W$7, SSIM $H \times W$8, and LPIPS $H \times W$9 (Kim, 21 Apr 2026). The same study also reports substantially better UV utilization: non-empty UV ratio rises from $K$0 to $K$1, collision rate falls from $K$2 to $K$3, and Gaussian retention increases from $K$4 to $K$5 (Kim, 21 Apr 2026). A common misconception is that increasing UV resolution or per-slot capacity $K$6 is sufficient; OT-UVGS explicitly states that increasing UV resolution or $K$7 does not guarantee higher effective representation under spherical projection (Kim, 21 Apr 2026).

3. Surface attachment, deformation, and Gaussian textures

A stricter form of mesh-aware Gaussian UV mapping uses the mesh not merely as a coordinate chart but as the geometric support of the Gaussian representation. UV Gaussians is exemplary: it starts from a SMPL-X template in canonical T-pose, applies linear blend skinning to obtain a coarse posed mesh, and uses a Mesh U-Net in UV space to predict a per-vertex offset map for fine geometry (Jiang et al., 2024). The refined mesh $K$8 provides the 3D anchor positions for the Gaussian textures, while a Gaussian U-Net predicts position residual, scale, rotation, opacity, and color in UV space:

$K$9

Every valid texel in the UV mask corresponds to a Gaussian in 3D, which yields dense, uniform coverage and stable semantic correspondence across poses (Jiang et al., 2024). The reported system typically uses $C$0190,000 Gaussians for a $C$1 UV map and achieves average novel-view PSNR $C$2, SSIM $C$3, and LPIPS $C$4; for novel pose synthesis it reports PSNR $C$5, SSIM $C$6, and LPIPS $C$7 (Jiang et al., 2024).

MeGA adopts a hybrid strategy. The face is modeled with an enhanced FLAME mesh and a UV displacement map $C$8 that refines per-vertex geometry, while hair is modeled by 3D Gaussian Splatting and animated with rigid alignment plus an MLP-based deformation field (Wang et al., 2024). The facial branch relies on a shared UV parametrization for both geometry refinement and neural texture application, and the final rendering uses occlusion-aware blending between the mesh-based head and Gaussian-based hair:

$C$9

On NeRSemble, MeGA reports PSNR $(x_i,y_i,z_i)$0, SSIM $(x_i,y_i,z_i)$1, and LPIPS $(x_i,y_i,z_i)$2, compared with GaussianAvatars at PSNR $(x_i,y_i,z_i)$3, SSIM $(x_i,y_i,z_i)$4, and LPIPS $(x_i,y_i,z_i)$5 (Wang et al., 2024).

3D Gaussian Model extends the idea by binding 3D Gaussians directly in a UV-parameterized texture space associated with a proxy mesh (Wang et al., 2024). Each Gaussian is represented in $(x_i,y_i,z_i)$6 coordinates and reprojected to world space through implicit shell mapping:

$(x_i,y_i,z_i)$7

This is not merely texture lookup on a mesh; the Gaussians themselves are UV-bound volumetric primitives, so animating the proxy mesh deforms the splats (Wang et al., 2024).

4. Continuous UV mappings and local Gaussian patches

Several methods replace discrete per-Gaussian UV assignment with continuous UV maps defined over surface points or ray–Gaussian intersections. Texture-GS disentangles geometry and appearance by learning a UV mapping MLP $(x_i,y_i,z_i)$8 together with an inverse map and a learnable texture (Xu et al., 2024). The UV objective combines cycle consistency in 3D, Chamfer-distance uniformity, and cycle consistency in 2D:

$(x_i,y_i,z_i)$9

At render time, Texture-GS does not assign one UV coordinate to each Gaussian center only. Instead, for each ray–Gaussian intersection it computes a local shading point

$$\rho_i = \sqrt{x_i^2 + y_i^2 + z_i^2},\quad \theta_i = \tan^{-1}(y_i, x_i),\quad \phi_i = \cos^{-1}\left(\frac{z_i}{\rho_i}\right),$$0

and approximates its UV coordinate with a first-order Taylor expansion at the Gaussian center:

$$\rho_i = \sqrt{x_i^2 + y_i^2 + z_i^2},\quad \theta_i = \tan^{-1}(y_i, x_i),\quad \phi_i = \cos^{-1}\left(\frac{z_i}{\rho_i}\right),$$1

This yields smoothly varying UVs inside each splat footprint and enables real-time rendering on a single RTX 2080 Ti GPU (Xu et al., 2024).

GTAvatar similarly bridges Gaussian splatting and UV texturing by embedding each canonical Gaussian primitive’s local frame into a patch in the UV space of a FLAME mesh (Baert et al., 9 Dec 2025). For a ray–splat intersection with local coordinates $$\rho_i = \sqrt{x_i^2 + y_i^2 + z_i^2},\quad \theta_i = \tan^{-1}(y_i, x_i),\quad \phi_i = \cos^{-1}\left(\frac{z_i}{\rho_i}\right),$$2, the mapping to UV is written as

$$\rho_i = \sqrt{x_i^2 + y_i^2 + z_i^2},\quad \theta_i = \tan^{-1}(y_i, x_i),\quad \phi_i = \cos^{-1}\left(\frac{z_i}{\rho_i}\right),$$3

which reduces the local-frame-to-UV conversion to a single matrix multiplication after precomputation (Baert et al., 9 Dec 2025). The resulting global UV textures store albedo, roughness, specular, and normals, and the system uses a Cook-Torrance BRDF with split-sum approximation for relighting (Baert et al., 9 Dec 2025).

More general neural UV mapping work provides adjacent principles rather than Gaussian-specific pipelines. Nuvo represents UV mapping with chart assignment, texture-coordinate, and surface-coordinate MLPs, optimized by cycle consistency, entropy, Chamfer, chart clustering, and distortion regularization for visible points on unruly geometry (Srinivasan et al., 2023). AUV-Net learns aligned UV maps across shapes through a UV mapper, masker, basis generators, and a texture-alignment module inspired by linear subspace learning, while using smoothness and cycle-style losses to preserve locality and encourage injective mappings (Chen et al., 2022). These methods indicate that Gaussian UV mapping can be continuous, charted, and semantically aligned without depending strictly on classical vertex-based parameterization.

5. Atlas construction, video coding, relighting, and editability

For dynamic scenes, Gaussian UV mapping has been used as a storage and streaming format rather than only as a rendering aid. PackUV maps all Gaussian attributes into a sequence of structured, multi-scale UV atlas and directly optimizes Gaussian parameters in the UV domain (Rai et al., 26 Feb 2026). It allocates up to $$\rho_i = \sqrt{x_i^2 + y_i^2 + z_i^2},\quad \theta_i = \tan^{-1}(y_i, x_i),\quad \phi_i = \cos^{-1}\left(\frac{z_i}{\rho_i}\right),$$4 layers per UV pixel ordered by opacity,

$$\rho_i = \sqrt{x_i^2 + y_i^2 + z_i^2},\quad \theta_i = \tan^{-1}(y_i, x_i),\quad \phi_i = \cos^{-1}\left(\frac{z_i}{\rho_i}\right),$$5

and stores deeper layers at progressively lower spatial resolutions:

$$\rho_i = \sqrt{x_i^2 + y_i^2 + z_i^2},\quad \theta_i = \tan^{-1}(y_i, x_i),\quad \phi_i = \cos^{-1}\left(\frac{z_i}{\rho_i}\right),$$6

The layers are then packed into a single 2D atlas with approximately $$\rho_i = \sqrt{x_i^2 + y_i^2 + z_i^2},\quad \theta_i = \tan^{-1}(y_i, x_i),\quad \phi_i = \cos^{-1}\left(\frac{z_i}{\rho_i}\right),$$7 packing efficiency, compared with approximately $$\rho_i = \sqrt{x_i^2 + y_i^2 + z_i^2},\quad \theta_i = \tan^{-1}(y_i, x_i),\quad \phi_i = \cos^{-1}\left(\frac{z_i}{\rho_i}\right),$$8 for a naïve grid (Rai et al., 26 Feb 2026).

PackUV couples this atlas with flow-guided Gaussian labeling and video keyframing, using forward RAFT optical flow to identify dynamic regions and freezing static Gaussian gradients during non-keyframe intervals (Rai et al., 26 Feb 2026). It further adopts Low Precision Optimization: all attributes except Gaussian position are quantized to 8 bits during training, position uses 16-bit quantization or two 8-bit channels, and a Straight-Through Estimator preserves gradients (Rai et al., 26 Feb 2026). The resulting atlas sequence is compatible with standard video codecs such as FFV1, HuffYUV, and HEVC in lossless mode, enabling bit-exact reconstruction with lossless coding and scaling to sequences up to 30 minutes (Rai et al., 26 Feb 2026). The same work introduces PackUV-2B, described as the largest multi-view video dataset to date, with more than 50 synchronized cameras, 100 sequences, and 2B frames (Rai et al., 26 Feb 2026).

Editability is a second major application. Texture-GS stores view-independent appearance in a 2D texture map, so texture swapping and painting become mesh-style operations on $$\rho_i = \sqrt{x_i^2 + y_i^2 + z_i^2},\quad \theta_i = \tan^{-1}(y_i, x_i),\quad \phi_i = \cos^{-1}\left(\frac{z_i}{\rho_i}\right),$$9 (Xu et al., 2024). GTAvatar pushes this further by reconstructing continuous editable material head textures and enabling relighting without additional optimization (Baert et al., 9 Dec 2025). MeGA supports hairstyle alteration and texture editing because the mesh and Gaussian components are independently controlled but coherently blended (Wang et al., 2024). DeMapGS extracts diffuse, normal, and displacement maps from surface-attached splats so that the reconstructed mesh can inherit Gaussian rendering quality in conventional rendering pipelines (Zhou et al., 11 Dec 2025).

6. Evaluation, misconceptions, and open directions

Evaluation of mesh-aware Gaussian UV mapping inherits criteria from both Gaussian rendering and classical parameterization. On the rendering side, PSNR, SSIM, and LPIPS dominate, as seen in OT-UVGS, UV Gaussians, MeGA, and GTAvatar (Kim, 21 Apr 2026, Jiang et al., 2024, Wang et al., 2024, Baert et al., 9 Dec 2025). On the parameterization side, the benchmark in “A Dataset and Benchmark for Mesh Parameterization” provides maximum area distortion, average area discrepancy, minimum and maximum singular values, percentage of flipped triangles, maximum angle distortion, average angle discrepancy, symmetric Dirichlet energy, resolution, artist correlation, remeshed, and seam statistics across 337 3D models gathered from Blenderkit, Sketchfab, Thingiverse, PolyHaven, and Keenan Crane’s model repository (Shay et al., 2022). A plausible implication is that this benchmark provides an evaluation template for learned or Gaussian-aware UV methods whenever they output a mesh-aligned atlas rather than only an image-like tensor.

Two misconceptions recur in the literature. First, UV mapping is not synonymous with mesh awareness. OT-UVGS improves UV utilization and quality through globally coupled rank-based allocation but “does not explicitly use mesh structure or topology in its mapping step” (Kim, 21 Apr 2026). Second, low geometric distortion is not the only practical objective. Artist-provided UV maps may contain injectivity violations or deliberate distortion, and most automatic methods have low correspondence in per-triangle distortion to artist UV maps (Shay et al., 2022). Recent unsupervised mesh parameterization work therefore augments geometry-preserving objectives with semantic and visibility-aware terms, using Shape Diameter Function segmentation and an ambient-occlusion-weighted seam loss to steer seams toward occluded regions (Zamani et al., 29 Sep 2025). This suggests that future mesh-aware Gaussian UV mapping may increasingly optimize for editability, semantic grouping, and seam visibility in addition to reconstruction fidelity.

The present field spans a broad design space. At one end are structured but topology-agnostic UV tensors such as UVGS and OT-UVGS; at the other are surface-attached or patch-based systems such as UV Gaussians, GTAvatar, DeMapGS, and 3DGM, in which UV coordinates are inseparable from mesh deformation, relighting, or map extraction (Rai et al., 3 Feb 2025, Kim, 21 Apr 2026, Jiang et al., 2024, Baert et al., 9 Dec 2025, Zhou et al., 11 Dec 2025, Wang et al., 2024). Between them lie continuous neural parameterizations and atlas-packing systems that treat UV space as a learned or codec-native computational domain (Xu et al., 2024, Srinivasan et al., 2023, Rai et al., 26 Feb 2026). Taken together, these works define mesh-aware Gaussian UV mapping not as a single algorithmic pattern, but as the convergence of Gaussian splatting, UV parameterization, and surface-aware representation design.