NVIDIA Meshtron Overview

What is Meshtron?

Meshtron is an autoregressive model for generating high-quality 3D meshes with artist-like topology. It provides a scalable, data-driven solution capable of generating intricate meshes with up to 64K faces at 1024-level coordinate resolution.

Key Features

• Scalability: Over 10x higher face count than existing methods
• Resolution: 8x higher coordinate resolution compared to previous approaches
• Efficiency: 2.5x faster token throughput and 50% memory savings
• Control Options: Adjustable parameters for customization

Control Inputs

• Point Cloud: Determines output mesh shape
• Face Count: Controls mesh density
• Quad Ratio: Switches between quad and triangle tessellation
• Creativity: Adjusts generation of additional details

Technical Architecture

Meshtron uses an Hourglass Transformer architecture with three stages:

1. Coordinate level: Processes every token
2. Vertex level: 3x reduction in sequence length
3. Face level: 9x reduction in sequence length

Key Innovations

• Sliding window attention with 8192 face context window
• Cross-attention implementation for control inputs
• Efficient token sequence processing
• Bottom-to-top generation process

Applications

• Standalone remesher for existing meshes
• Integration with text-to-3D systems
• Integration with image-to-3D models
• Generation of artist-grade meshes