Mitigating Seam Artifacts in 3D Texture Generation

159 concepts113 papers2 proposals

In 3D computer graphics, applying a 2D image to a 3D surface is traditionally achieved through UV mapping—a process akin to cutting open a complex 3D shape and flattening it out so it can lie on a 2D plane. While this pipeline is a cornerstone of standard graphics hardware, it introduces a fundamental mathematical conflict when paired with modern generative AI models.

Because complex 3D topologies cannot be flattened without stretching or tearing, the UV unwrapping process forces cuts across the surface. These cuts create UV seams: artificial boundaries where points that are physically adjacent in 3D space end up separated into distant "islands" on the 2D texture map. When 2D generative models (such as diffusion models or GANs) operate on these flat images, they lack the spatial awareness required to maintain continuity across these boundaries, resulting in visible, jagged seam artifacts on the final 3D object.

An emerging research direction focuses on resolving these discontinuities, bridging the gap between 2D generative efficiency and 3D spatial coherence.


The Root Cause: Topological Disruption

To understand why traditional texturing pipelines fail in automated workflows, we must look at how standard neural networks process data. Convolutional layers and attention mechanisms operate on the assumption of spatial locality; they assume that pixels next to each other in an image represent features that are next to each other in reality.

As Foti et al. observe, wrapping a 2D texture onto complex 3D surfaces inevitably introduces geometric distortions and visible seams because a continuous 3D shape cannot be mapped to a flat surface without tearing. When a mesh is cut into a 2D layout, this spatial adjacency is broken.

As a result, a generative model treating a UV map as an ordinary image remains completely blind to the 3D geometry. At the edge of a UV island, the network's local filters process empty padding or an unrelated patch of the mesh. When the texture is wrapped back onto the 3D model, the patterns, colors, and high-frequency details fail to align at the cut edges. Liu et al. emphasize that this loss of global context frequently leads to visual seams and inconsistent texture alignment, posing a severe constraint for automated asset generation.

3D Space:  [ Point A ] <--- adjacent ---> [ Point B ]
                 \                             /
UV Cut:           \                           /
                   v                         v
2D Atlas:  [ Island 1: Point A ]  ...  [ Island 2: Point B ]
           (Processed independently by 2D convolutions)

Strategy 1: Injecting 3D Proximity into 2D Synthesis

One prominent strategy to eliminate seam artifacts without abandoning the industry-standard UV map is to inject 3D geometric awareness directly into the 2D generative pipeline. If the network knows which distant 2D pixels are actually 3D neighbors, it can enforce continuity.

3D Spatial Coordinate Conditioning

Zeng and Zhao address this by incorporating explicit 3D spatial coordinates into the UV mapping and generation process. By conditioning the generative model on the underlying 3D coordinates of each texel, the network learns to recognize that geographically distant patches in the 2D plane are adjacent on the 3D surface. This spatial anchoring prevents the disjointed patterns that typically occur when 2D diffusion models operate blindly on fragmented UV layouts.

Cross-Dimensional Feature Blending

Rather than just conditioning the model, texturing frameworks can actively smooth features across known boundaries. Cheng and Mu demonstrate a framework that performs texture inpainting and refinement by identifying UV boundaries and blending colors based on their 3D spatial neighborhood rather than 2D coordinate distance. By prioritizing geometric proximity over flat image coordinates, the system ensures smooth color transitions and visual continuity across seams, even when the underlying UV map is highly fragmented or poorly unwrapped.

Hybrid 2D-3D Architectures

Instead of choosing between 2D efficiency and 3D continuity, some approaches combine both natively within the network architecture. Yu and Yu introduce a model that interleaves 2D convolutions on UV maps with 3D attention mechanisms on point clouds.

$$\text{Feature Update} = \text{Conv2D}(\text{UV Map}) + \text{Attention3D}(\text{Point Cloud})$$

This hybrid approach leverages the computational speed of 2D convolutions for generating high-resolution details while using 3D attention to bridge the topological gaps caused by UV cuts, allowing features to flow naturally across seams during the generation pass.


Strategy 2: Continuous Neural UV Mapping

An alternative to fixing artifacts post-hoc is to rethink how the UV map itself is constructed. Traditional UV unwrapping tools operate discretely on mesh topology, often slicing complex or noisy AI-generated models into thousands of tiny, disconnected texture charts.

To solve this fragmentation, Srinivasan et al. propose a neural field-based approach called Nuvo. Instead of relying on discrete mesh operations, they represent UV mappings as continuous functions optimized via Multi-Layer Perceptrons (MLPs).

By optimizing the UV map based on visible surface points rather than the raw underlying mesh topology, the mapping process ignores internal or occluded geometry. The continuous nature of the neural network inherently smooths out the coordinate space, preventing the fragmentation of texture atlases and eliminating the harsh, multi-island boundaries that cause traditional seam artifacts.


Strategy 3: UV-Free Surface Generation

The most radical and effective way to eliminate UV seam artifacts is to bypass the 2D intermediate representation entirely, generating textures directly on the 3D manifold.

Traditional:  [3D Mesh] -> [2D UV Unwrapping] -> [2D Diffusion] -> [Seam Artifacts]
UV-Free:      [3D Mesh] -> [Direct Surface Diffusion / Heat Flow] -> [Seamless Texture]

Geodesic Heat Diffusion

Foti et al. introduce a framework (UV3-TeD) that skips UV mapping altogether by combining denoising diffusion with the physics of heat flow. Instead of diffusing noise across a flat 2D pixel grid, their method diffuses features directly across the continuous 3D surface using geodesic heat diffusion. Because the generation process operates natively on the 3D shape, there are mathematically no cuts, charts, or boundaries, completely avoiding geometric distortions and seams.

Octree-Based Spatial Diffusion

Operating directly on 3D structures can be computationally expensive. To make direct surface texturing viable, Liu et al. utilize an octree-based diffusion model (TexOct). By representing dense point clouds as efficient, hierarchical octrees, the model synthesizes textures directly on the surface of the 3D model. This spatial structuring bypasses the pitfalls of UV parameterization entirely, ensuring that the generated texture remains topologically consistent and completely immune to the stretching and seam discontinuities that plague traditional graphics pipelines.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 116 papers
  2. Version 1 · 2026-10-04 · 125 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_texture_seam_artifacts_b026525a,
  title = {texture seam artifacts},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/texture-seam-artifacts/rdr_b026525a5dc34dd797252b1be25ef32d},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 2, frozen text):

@misc{lacuna_texture_seam_artifacts_b026525a_v2,
  title = {texture seam artifacts},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/texture-seam-artifacts/rdr_b026525a5dc34dd797252b1be25ef32d/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake