TopGen: Learning Structural Layouts and Cross-Fields for Quadrilateral Mesh Generation

TopGen is a robust, learning-based framework that simultaneously predicts structural layouts and cross-fields to generate high-quality quadrilateral meshes, overcoming the efficiency and topological limitations of traditional methods while preserving geometric integrity and editability.

The Gist

Imagine you have a beautiful, intricate 3D sculpture made of thousands of tiny, jagged triangles (like a low-poly video game character or a 3D-printed statue). While this shape looks good, it's a nightmare for animators and designers. To make it move, bend, or get painted, they need to cover it with a neat, organized grid of squares (quadrilaterals), much like laying down a perfect floor of tiles.

The problem? Doing this manually is like trying to tile a bathroom floor while blindfolded—it takes artists days or weeks. Doing it automatically with old computer programs is like trying to tile that same floor with a sledgehammer: it's fast, but it often breaks the tiles, leaves gaps, or creates a messy, chaotic pattern that doesn't make sense.

Enter TopGen: The "Smart Architect" for 3D Shapes.

This paper introduces TopGen, a new AI system that acts like a master architect who can instantly look at a messy pile of triangles and say, "I know exactly how to lay the tiles."

Here is how it works, using some simple analogies:

1. The Old Way vs. The New Way

• The Old Optimization Way: Imagine trying to solve a massive, complex math puzzle where you have to move every single tile at the same time to make them fit perfectly. If the puzzle is too big (a high-resolution model), the computer gets stuck, crashes, or takes hours to finish.
• The Old Learning Way: Newer AI methods tried to guess the pattern by looking at the "wind direction" (called a cross-field) on the surface. It's like telling a tile layer, "Just follow the breeze." The problem? The breeze doesn't know where the sharp corners of a table are. The result? The tiles look smooth in the middle but get jagged and broken at the edges.
• The TopGen Way: TopGen does two things at once, just like a human artist does.
  1. It draws the "Skeleton": It first identifies the most important lines (the sharp edges, the corners, the spine of a character). Think of this as drawing the blueprint or the frame of a house before you start building.
  2. It maps the "Flow": While drawing the skeleton, it also figures out the direction the tiles should flow in the empty spaces between the lines.

2. How TopGen "Sees" the Shape

Instead of looking at the shape as a rigid mesh of connected triangles, TopGen treats it like a cloud of dust.

• The "Sharp Edge" Trick: Imagine sprinkling glitter on a statue. Most glitter falls evenly, but if you shake the statue, extra glitter clumps onto the sharp edges and corners. TopGen does this digitally. It grabs extra data points from the sharp corners (so it doesn't miss details) and spreads them out evenly elsewhere.
• The "Dual-Query" Brain: This is the secret sauce. TopGen asks two different questions simultaneously:
  • The Edge Question: "Is this specific line part of the main skeleton?" (Yes/No).
  • The Face Question: "Which way should the tiles flow in this little patch of surface?" (Direction).
    By asking these two questions at the exact same time, the AI learns that the "skeleton" and the "flow" are best friends. You can't have one without the other.

3. The "TopGen-220K" Dataset

To teach this AI, the researchers couldn't just use random internet shapes. They needed a massive library of "perfect examples."

• They built a dataset called TopGen-220K.
• Imagine a library with 220,000 books. But instead of text, each book contains:
  1. The messy original shape.
  2. The perfect "skeleton" drawn by a human expert.
  3. The perfect "flow" map.
  4. The final, beautiful square grid.
• This is the first time such a massive, high-quality library has existed. It's like giving the AI a million hours of apprenticeship under the world's best 3D artists.

4. Why It Matters

• Speed: While old methods might take minutes or hours to process a complex character, TopGen does it in less than a second. It's like going from hand-crafting a tile floor to using a 3D printer that lays it down instantly.
• Quality: It doesn't just make squares; it makes smart squares. It respects the sharp edges of a robot's arm or the curves of a character's face. It doesn't break the geometry.
• Robustness: It works even on "messy" shapes that other computers can't handle (like AI-generated models that are full of holes or weird connections).

The Bottom Line

TopGen is a breakthrough because it stops trying to force a mathematical solution onto a 3D shape. Instead, it learns the "art" of layout. It understands that to build a good grid, you first need to understand the structure, and then you can fill in the details.

It's the difference between a robot blindly stamping squares onto a wall and a master mason who looks at the wall, sees the cracks and corners, and lays the bricks perfectly to make a beautiful, sturdy structure.

Technical Summary

Here is a detailed technical summary of the paper "TopGen: Learning Structural Layouts and Cross-Fields for Quadrilateral Mesh Generation."

1. Problem Statement

High-quality quadrilateral (quad) mesh generation is a critical bottleneck in computer graphics, essential for animation, industrial design, and digital content creation. While 3D generative AI can produce high-resolution models, the resulting topologies are often fragmented and incompatible with professional pipelines.

The paper identifies two main limitations in existing approaches:

• Optimization-Based Methods: Traditional algorithms (e.g., Instant-Meshes, QuadriFlow) rely on global optimization and cross-field guidance. They suffer from severe efficiency bottlenecks (slow processing on high-res models), require high-quality manifold inputs, and often fail to capture high-level semantic structures, leading to irrational singularity distributions and distorted edge flows.
• Learning-Based Methods: Emerging deep learning approaches focus primarily on predicting cross-fields (orientation fields). However, relying solely on these "soft constraints" often results in the loss of critical geometric details (sharp edges/corners), jagged boundaries, and a lack of editability. They fail to explicitly model the "structural skeleton" that professional artists manually create.

2. Methodology: TopGen Framework

TopGen is a learning-based framework that mimics the professional "layout-first, generation-second" workflow. It simultaneously predicts structural layouts (hard constraints) and cross-fields (soft constraints) to guide quad mesh generation.

A. Data Representation & Encoder

• Input Strategy: Instead of processing raw meshes directly, TopGen uses point cloud sampling. It employs a Sharp Edge Sampling (SES) strategy to sample points along sharp features (high dihedral angles) and uniformly across the surface. This ensures robustness to non-manifold geometries and low-quality initial topologies.
• Geometry-Aware Encoder: Utilizes a pre-trained shape encoder (based on Dora-VAE) with a Dual Cross-Attention (DCA) mechanism.
  • It processes salient points (features) and uniform points (surface) separately.
  • It fuses these features via cross-attention and self-attention to create a continuous latent space ($Z$) that captures both global shape semantics and fine-grained local geometry.

B. Dual-Query Decoder

The core innovation is a Dual-Query Decoder that performs parallel tasks using two types of query points derived from the input mesh topology:

1. Edge Queries ($Q_e$): Placed at edge midpoints.
   • Task: Structural Line Classification (Binary).
   • Mechanism: Aggregates local topological context (vertices and adjacent face midpoints) via self-attention. It predicts the probability of an edge belonging to the structural layout. A bias term is added for salient edges to enforce sharp feature alignment.
2. Face Queries ($Q_f$): Placed at face centroids.
   • Task: Cross-Field Regression.
   • Mechanism: Aggregates adjacent face contexts. It predicts a 2-RoSy cross-field (represented as a polyvector) to guide internal edge orientation.
   • Weighting: Queries near salient edges are weighted higher to ensure the field aligns with sharp features.

C. Remeshing Pipeline

Once the structural lines and cross-fields are predicted:

• Hard Constraints: Predicted structural lines are thresholded to form polylines, acting as rigid boundaries to prevent feature drift.
• Soft Constraints: The predicted cross-fields guide the parameterization and field-aligned mesh generation.
• Extraction: Integer iso-lines are traced in the parameter domain to extract the final quad mesh, ensuring high regularity and low distortion.

3. Key Contributions

1. TopGen Framework: The first learning-based method to jointly predict structural layouts and cross-fields. This dual approach ensures both the integrity of external geometric boundaries (via structural lines) and the regularity of internal edge flows (via cross-fields).
2. Dual-Query Decoder: A novel architecture that leverages the synergy between structure and field by performing parallel edge classification and field regression, explicitly preserving mesh topology during decoding.
3. TopGen-220K Dataset: A large-scale, comprehensive dataset containing 220,000 curated samples. It provides paired data of raw triangular meshes, high-quality structural layouts, cross-fields, and corresponding quad meshes. This fills a critical gap in the availability of structural feature data for training.

4. Experimental Results

TopGen was evaluated against state-of-the-art methods (Instant-Meshes, QuadriFlow, QuadWild, NeurCross) on diverse datasets (QuadWild-300, AI-generated models).

• Geometric Fidelity: TopGen achieved the lowest Chamfer Distance (CD) across all test cases, indicating superior preservation of the original geometry compared to competitors.
• Topological Quality: It produced meshes with fewer singularities and lower average Scaled Jacobian values (indicating better element quality) without structural corruption.
• Robustness: Unlike optimization-based methods that fail on non-manifold or AI-generated meshes, TopGen handled these inputs successfully, producing clean, game-ready topologies.
• Efficiency:
  • Speed: TopGen infers in <1 second (average ~0.03s), whereas optimization-based methods (like NeurCross or QuadWild) take hundreds to thousands of seconds.
  • Stability: It avoided the numerical instability (NaN values) observed in NeurCross when processing AI-generated models.
• Ablation Studies: Experiments confirmed that using only cross-fields leads to jagged boundaries, while using only structural lines results in disorganized internal flows. The joint prediction is essential for high-quality results.

5. Significance

• Bridging the Gap: TopGen successfully bridges the gap between the speed of learning-based methods and the topological quality of manual modeling.
• Industry Applicability: By generating meshes that adhere to professional "layout-first" standards, TopGen enables the integration of AI-generated 3D content into production pipelines (games, VFX) without requiring extensive manual retopology.
• Data Contribution: The release of TopGen-220K provides a foundational resource for future research in geometric deep learning, specifically for tasks requiring structural understanding.
• Future Potential: The architecture shows promise for other layout-prediction tasks such as UV unwrapping and part segmentation.