README.md

---
dataset_info:
  features:
  - name: voxels_occupancy
    dtype:
      array4_d:
        shape:
        - 1
        - 32
        - 32
        - 32
        dtype: float32
  - name: voxels_colors
    dtype:
      array4_d:
        shape:
        - 3
        - 32
        - 32
        - 32
        dtype: float32
  - name: model_id
    dtype: string
  - name: name
    dtype: string
  - name: categories
    sequence: string
  - name: tags
    sequence: string
  - name: metadata
    struct:
    - name: is_augmented
      dtype: bool
    - name: original_file
      dtype: string
    - name: num_occupied
      dtype: int32
    - name: split
      dtype: string
  splits:
  - name: train
    num_bytes: 207426028571
    num_examples: 515177
  - name: test
    num_bytes: 10942650293
    num_examples: 27115
  download_size: 4338398625
  dataset_size: 218368678864
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: test
    path: data/test-*
---

# BlockGen-3D Dataset

## Overview

BlockGen-3D is a large-scale dataset of voxelized 3D models with accompanying text descriptions, specifically designed for text-to-3D generation tasks. By processing and voxelizing models from the [Objaverse dataset](https://huggingface.co/datasets/allenai/objaverse), we have created a standardized representation that is particularly suitable for training 3D diffusion models.

Our dataset provides two types of representations: shape-only models represented as binary occupancy grids, and colored models with full RGBA information. Each model is represented in a 32×32×32 voxel grid. 
The dataset contains 542,292 total samples, with:
- 515,177 training samples
- 27,115 test samples

## Data Format and Structure

Every sample in our dataset consists of a voxel grid with accompanying metadata. The voxel grid uses a consistent 32³ resolution, with values structured as follows:

For shape-only data:
- `voxels_occupancy`: A binary grid of shape [1, 32, 32, 32] where each voxel is either empty (0) or occupied (1)

For colored data:
- `voxels_colors`: RGB color information with shape [3, 32, 32, 32], normalized to [0, 1]
- `voxels_occupancy`: The occupancy mask as described above

Each sample also includes rich metadata:
- Text descriptions derived from the original Objaverse dataset
- Categorization and tagging information
- Augmentation status and original file information
- Number of occupied voxels for quick filtering or analysis

## Using the Dataset

### Basic Loading and Inspection

```python
from datasets import load_dataset

# Load the dataset
dataset = load_dataset("PeterAM4/blockgen-3d")

# Access splits
train_dataset = dataset["train"]
test_dataset = dataset["test"]

# Examine a sample
sample = train_dataset[0]
print(f"Sample description: {sample['name']}")
print(f"Number of occupied voxels: {sample['metadata']['num_occupied']}")
```

### Working with Voxel Data

When working with the voxel data, you'll often want to handle both shape-only and colored samples uniformly. Here's a utility function that helps standardize the processing:

```python
import torch

def process_voxel_data(sample, default_color=[0.5, 0.5, 0.5]):
    """Process voxel data with default colors for shape-only data.
    
    Args:
        sample: Dataset sample
        default_color: Default RGB values for shape-only models (default: gray)
    
    Returns:
        torch.Tensor: RGBA data with shape [4, 32, 32, 32]
    """
    occupancy = torch.from_numpy(sample['voxels_occupancy'])
    
    if sample['voxels_colors'] is not None:
        # Use provided colors for RGBA samples
        colors = torch.from_numpy(sample['voxels_colors'])
    else:
        # Apply default color to shape-only samples
        default_color = torch.tensor(default_color)[:, None, None, None]
        colors = default_color.repeat(1, 32, 32, 32) * occupancy
    
    # Combine into RGBA format
    rgba = torch.cat([colors, occupancy], dim=0)
    return rgba
```

### Batch Processing with DataLoader

For training deep learning models, you'll likely want to use PyTorch's DataLoader. Here's how to set it up with a simple prompt strategy:

For basic text-to-3D generation, you can use the model names as prompts:
```python
def collate_fn(batch):
    """Simple collate function using basic prompts."""
    return {
        'voxels': torch.stack([process_voxel_data(x) for x in batch]),
        'prompt': [x['name'] for x in batch]  # Using name field as prompt, could also use more complex prompts from categories and tags dict keys
    }

# Create DataLoader
dataloader = DataLoader(
    train_dataset,
    batch_size=32,
    shuffle=True,
    collate_fn=collate_fn
)
```

## Visualization Examples

![Dataset Usage Example](https://huggingface.co/datasets/PeterAM4/blockgen-3d/resolve/main/comparison_wooden.gif)
![Dataset Usage Example](https://huggingface.co/datasets/PeterAM4/blockgen-3d/resolve/main/comparison_char.gif)
*Dataset Usage Example - Showing voxelization process*

## Dataset Creation Process

Our dataset was created through the following steps:

1. Starting with the Objaverse dataset
2. Voxelizing 3D models at 32³ resolution
3. Preserving color information where available
4. Generating data augmentations through rotations
5. Creating train/test splits (95%/5% split ratio)

## Training tips

- For shape-only models, use only the occupancy channel
- For colored models:

  - Apply colors only to occupied voxels
  - Use default colors for shape-only samples

## Citation

If you use this dataset in your research, please cite:

```
@misc{blockgen2024,
  title={BlockGen-3D: A Large-Scale Dataset for Text-to-3D Voxel Generation},
  author={Peter A. Massih},
  year={2024},
  publisher={Hugging Face}
}
```

Please also cite the original Objaverse dataset:

```
@article{objaverse2023,
  title={Objaverse: A Universe of Annotated 3D Objects},
  author={Matt Deitke and Dustin Schwenk and Jordi Salvador and Luca Weihs and Oscar Michel and Eli VanderBilt and Ludwig Schmidt and Kiana Ehsani and Aniruddha Kembhavi and Ali Farhadi},
  journal={arXiv preprint arXiv:2304.02643},
  year={2023}
}
```

## Limitations

- Fixed 32³ resolution limits fine detail representation
- Not all models have color information
- Augmentations are limited to rotations
- Voxelization may lose some geometric details