Research map / 3D vision / Face and shape

Mesh BDF: Barycentric Dominance Field for 3D Native Mesh Generation

NeurIPS 2026 · 3D vision › Face and shape · approach: Textured mesh generation · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes the Barycentric Dominance Field (BDF), a scalar field on triangle meshes, and integrates it as a texture-like signal into the Trellis.2 sparse-voxel VAE and flow pipeline to generate native meshes that beat AR baselines on Toys4k.

Problem

Autoregressive native mesh generators are limited in face count, vertex resolution, texture support, and robustness to token errors, while diffusion models need fixed-length continuous representations of meshes.

Approach

Encode mesh connectivity as a continuous scalar field (max barycentric coordinate) stored on dual-contouring voxel edges, so a standard texture-style VAE and diffusion transformer can generate it; mesh topology is recovered by a BDF-guided flood-fill over the DC quad graph.

Key results

• On Toys4k 0~10k faces, the method reaches CD 0.102e-3, HD 0.038, NC 0.929, versus best AR baseline MeshRipple at CD 3.830e-3, HD 0.154, NC 0.895.
• On Toys4k 0~40k faces, CD 0.123e-3 versus MeshRipple 15.584e-3 and DeepMesh 25.113e-3.
• Ablation: decoupled VAE with BCE loss gives F-score 0.599 at 0.5V and 0.991 at 3V, better than MSE and jointed VAE.

Surprising findings

• MSE loss matches BCE on CD and HD but falls well short on strict F-score, indicating the sharp vertex peaks need BCE.

Limitations

• DC grid resolution loses topology of very small faces that miss grid intersections.
• Self-intersecting triangles in the original mesh can mislead the flood-fill and create erroneous connections.
• Evaluation uses a small Toys4k subset (200 and 100 meshes) and conditions AR baselines on point clouds sampled from Trellis.2 outputs.

Experimental setup

Datasets: Texverse Toys4k

Benchmarks: Toys4k

Metrics: Chamfer Distance Hausdorff Distance Normal Consistency F-score

Baselines: MeshSilksong DeepMesh MeshRipple

Builds on: Trellis.2 FlexiCubes Dual Contouring AdamW

Closest papers by meaning

• Single Mesh Diffusion Models with Field Latents for Texture Generation (CVPR 2024 · similarity 0.88)
• ARMesh: Autoregressive Mesh Generation via Next-Level-of-Detail Prediction (NeurIPS 2025 · similarity 0.87)
• CraftsMan3D: High-fidelity Mesh Generation with 3D Native Diffusion and Interactive Geometry Refiner (CVPR 2025 · similarity 0.86)
• SILSA: Sliding-Window Slice Latents for Topology-Preserving High-Resolution 3D Generation (NeurIPS 2026 · similarity 0.86)
• MeshGen: Generating PBR Textured Mesh with Render-Enhanced Auto-Encoder and Generative Data Augmentation (CVPR 2025 · similarity 0.86)
• Sin3DM: Learning a Diffusion Model from a Single 3D Textured Shape (ICLR 2024 · similarity 0.86)