Nexus: Native Mesh Generation with Diffusion

2026-07-15 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Gaming & Interactive Media, 3D Modeling & Computer Graphics · Depth: Expert, quick

Summary

Nexus is a novel diffusion method designed for generating high-quality triangle meshes, crucial for film, gaming, and interactive 3D applications. It addresses limitations of mainstream autoregressive methods by employing a holistic approach with decoupled vertex and topology generation. For vertices, Nexus treats them as sparse voxels within an octree, utilizing a diffusion model to generate them in a coarse-to-fine manner. To model topology, the method introduces "Spacetime Interval," an extension of Spacetime Distance, which encodes arbitrary edge and face topology into continuous per-vertex embeddings. A second diffusion model then generates these embeddings on the previously generated vertices, enabling global and efficient recovery of complex topology. Experiments on Objaverse and Toys4K datasets, alongside in-the-wild images, demonstrate Nexus's superior performance compared to autoregressive and two-stage baselines, confirmed by a blind user study showing strong perceptual preference from 3D practitioners.

Key takeaway

For 3D practitioners and ML engineers developing high-quality 3D assets for gaming or film, Nexus presents a compelling alternative to traditional autoregressive mesh generation. Its decoupled diffusion approach for vertices and topology, using "Spacetime Interval" embeddings, directly addresses issues like error accumulation and inefficient inference. You should explore integrating diffusion-based mesh generation techniques to achieve superior perceptual quality and more robust 3D models, potentially streamlining your asset creation pipelines.

Key insights

Nexus uses a decoupled diffusion approach for vertex and topology generation to create high-quality 3D meshes, outperforming sequential methods.

Principles

• Decouple vertex and topology generation.
• Encode topology into continuous per-vertex embeddings.
• Utilize diffusion models for coarse-to-fine generation.

Method

Generate vertices coarse-to-fine using a diffusion model on octree-organized sparse voxels. Then, use a diffusion model to generate Spacetime Interval embeddings on these vertices for topology recovery.

In practice

• Generate high-quality meshes for 3D assets.
• Improve efficiency in 3D content creation.
• Circumvent sequential mesh modeling issues.

Topics

• Mesh Generation
• Diffusion Models
• 3D Asset Creation
• Topology Encoding
• Objaverse Dataset
• Spacetime Interval

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• Native3D: End-to-End 3D Scene Generation via Unified Mesh-Texture Modeling and Semantic Alignment — Native3D generates 3D scenes end-to-end by unifying mesh-texture modeling and semantic alignment, avoiding 2D conversion issues.
• PixGS: Pixel-Space Diffusion for Direct 3D Gaussian Splat Generation — PixGS directly generates high-quality 3D Gaussian Splats using pixel-space diffusion, bypassing latent compression artifacts.
• ExMesh: EXplicit Mesh Reconstruction with Topology Adaptation — ExMesh directly optimizes explicit meshes by integrating differentiable optimization with adaptive topology and real-time UV maintenance.
• SQuadGen: Generating Simple Quad Layouts via Chart Distance Fields — SQuadGen uses diffusion models and Chart Distance Fields to generate simple, artist-friendly quad layouts for 3D shapes.
• Real2SAM2Real: Generative 3D Caches as Complementary Context for Video Diffusion — Real2SAM2Real uses editable 3D caches from lifting models to provide robust geometric scaffolds for Video Diffusion Models, improving control and consistency.
• FLUX3D: High-Fidelity 3D Gaussian Generation with Diffusion-Aligned Sparse Representation — FLUX3D enhances 3D Gaussian Splatting generation fidelity by resolving representation and cross-modal alignment bottlenecks.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.