L La M A Mesh G G U F

Introduction

LLaMA-Mesh-GGUF is a quantized version of the LLaMA-Mesh model, designed to unify 3D mesh generation with language models. This offers advantages such as leveraging spatial knowledge embedded in large language models (LLMs) and enabling conversational 3D generation. The model represents vertex coordinates and face definitions of 3D meshes as plain text, allowing direct integration with LLMs. This approach allows the model to generate 3D meshes from text prompts, produce interleaved text and 3D outputs, and understand 3D meshes.

Architecture

The model is based on the Transformer network architecture, specifically using Llama 3.1. It accepts text input in string format and outputs text in a similar format, with a maximum token length of 8k. The model is compatible with NVIDIA Ada microarchitecture and supports the Linux operating system.

Training

LLaMA-Mesh-GGUF is fine-tuned on a dataset derived from Objaverse, using a subset of 30,000 mesh data filtered by face count. The model is fine-tuned using 32 GPUs. The dataset is licensed under the ODC-By v1.0 license, and the training process leverages Pytorch as the inference engine.

Guide: Running Locally

1. Set Up Environment: Ensure you have Python and Pytorch installed. Set up a virtual environment to manage dependencies.
2. Clone Repository: Clone the model repository from Hugging Face.
3. Install Dependencies: Navigate to the cloned directory and install required packages using pip install -r requirements.txt.
4. Run Model: Use provided scripts to initialize and run the model.
5. Hardware Recommendations: It is recommended to use cloud GPUs such as NVIDIA A100 for optimal performance.

License

LLaMA-Mesh is subject to two licenses:

1. NSCLv1 License: Permits non-commercial use only.
2. Llama 3.1 Community License Agreement: Requires compliance with the terms for redistributing and using Llama 3.1 materials. Full details can be found in the LLAMA_LICENSE.txt file. Additionally, the model is built with Llama 3.1 technology, and attribution is required as per the community license agreement.