MeshXL is a family of generative foundation models for 3D mesh creation, targeting researchers and developers in 3D content generation. It leverages a Neural Coordinate Field (NeurCF) representation, enabling large-scale sequential modeling of 3D mesh data using LLM techniques, offering a novel approach to generating unstructured 3D assets.

How It Works

MeshXL utilizes a Neural Coordinate Field (NeurCF) representation, which combines explicit coordinates with implicit neural embeddings. This approach allows for the seamless application of modern LLM methodologies to the generation of unstructured 3D mesh data. The NeurCF representation is designed to be simple yet effective for large-scale sequential mesh modeling, facilitating efficient and high-quality mesh generation.

Quick Start & Requirements

• Environment Setup: Run bash set_env.sh.
• Pre-trained Weights: Download from HuggingFace for 125M, 350M, and 1.3B parameter models. Use git lfs clone for fast downloading.
• Generation: Execute bash scripts/sample-*.sh for unconditional generation or accelerate launch sample_t2m.py ... for text-to-mesh.
• Fine-tuning: Requires ShapeNet data downloaded from HuggingFace and uses accelerate launch main.py ... with specific configurations.
• Dependencies: Requires PyTorch, HuggingFace libraries, and accelerate. Mixed precision (bf16) is recommended for training/generation.

Highlighted Details

• Offers pre-trained models of varying sizes (125M, 350M, 1.3B parameters).
• Supports unconditional 3D mesh generation and partial mesh completion.
• Provides fine-tuned checkpoints for specific categories (chair, table, lamp, bench).
• Enables text-to-mesh generation using language annotations.

Maintenance & Community

• Project accepted to NeurIPS 2024.
• Active updates with new data and code releases on HuggingFace.
• Project page and Arxiv paper available for detailed information.

Licensing & Compatibility

• The repository does not explicitly state a license in the provided README. Compatibility for commercial use or closed-source linking is not specified.

Limitations & Caveats

• Data preparation for training and evaluation is noted as "Work in Progress."
• Gradio demos for text-to-mesh generation are actively being worked on.