Roblox Cube 3D Mesh Generation Code Release

Roblox has released Cube 3D, a generative AI system designed for 3D intelligence, as part of their open-source initiative. This release includes code and model weights for text-to-shape generation, enabling developers to create 3D assets programmatically. Here’s a breakdown of the key details:

Getting Started with Cube 3D

To begin using Cube 3D, follow these steps:

1. Clone the Repository: Clone the Cube 3D repository from GitHub using the following command:

   git clone https://github.com/Roblox/cube.git

2. Install Dependencies: Navigate to the cloned directory and install the required dependencies in a virtual environment:

   cd cube
   pip install -e .[meshlab]

   Note: [meshlab] is optional and can be omitted for better compatibility, but mesh simplification will be disabled.

3. Install CUDA (if needed): For Windows users, install the CUDA toolkit and PyTorch with CUDA support:

   pip install torch --index-url https://download.pytorch.org/whl/cu124 --force-reinstall

4. Download Model Weights: Download the pre-trained model weights from Hugging Face:

   huggingface-cli download Roblox/cube3d-v0.1 --local-dir ./model_weights

Generating 3D Meshes

To generate 3D models using the downloaded weights, run the following command:

python -m cube3d.generate \
  --gpt-ckpt-path model_weights/shape_gpt.safetensors \
  --shape-ckpt-path model_weights/shape_tokenizer.safetensors \
  --fast-inference \
  --prompt "Broad-winged flying red dragon, elongated, folded legs."

This will generate an .obj file in the specified output directory. You can also render a turntable GIF of the mesh using the --render-gif flag (requires Blender installed and accessible in your system’s PATH).

Shape Tokenization and De-tokenization

To tokenize a 3D shape and reconstruct it, use the following command:

python -m cube3d.vq_vae_encode_decode \
  --shape-ckpt-path model_weights/shape_tokenizer.safetensors \
  --mesh-path ./outputs/output.obj

This will tokenize the .obj file and print the tokenized representation, as well as export the reconstructed mesh.

Minimalist Python API

Cube 3D also provides a Python API for integration into your projects. Here’s an example:

import torch
import trimesh
from cube3d.inference.engine import Engine, EngineFast

# Load checkpoint
config_path = "cube3d/configs/open_model.yaml"
gpt_ckpt_path = "model_weights/shape_gpt.safetensors"
shape_ckpt_path = "model_weights/shape_tokenizer.safetensors"
engine_fast = EngineFast(
  config_path, gpt_ckpt_path, shape_ckpt_path, device=torch.device("cuda")
)

# Inference
input_prompt = "A pair of noise-canceling headphones"
mesh_v_f = engine_fast.t2s([input_prompt], use_kv_cache=True)

# Save output
vertices, faces = mesh_v_f[0][0], mesh_v_f[0][1]
trimesh.Trimesh(vertices=vertices, faces=faces).export("output.obj")

Hardware Requirements

Cube 3D has been tested on Nvidia H100, A100, and Geforce 3080 GPUs, as well as Apple Silicon M Chips. A GPU with at least 24GB of VRAM is recommended for optimal performance.

Citation

If you find Cube 3D useful, consider citing their technical report:

@article{roblox2025cube,
  title = {Cube: A Roblox View of 3D Intelligence},
  author = {Roblox, Foundation AI Team},
  journal = {arXiv preprint arXiv:2503.15475},
  year = {2025}
}

Additional Resources

For more details, visit the Cube 3D GitHub repository.