VQVDB is a deep learning-powered compressor for volumetric data stored in OpenVDB. It uses Vector Quantized Variational Autoencoders (VQ-VAE) to learn a compact latent space and achieves up to 32x compression of float voxel grids, nearly lossless at the visual level.

VQVDB is designed for GPU-accelerated decoding via CUDA but also support CPU encoding / decoding, enabling real-time decompression of large volumes, with native support for integration into Houdini.

Each .vqvdb file stores:

I'm very bad at programming and couldn't compile pyopenvdb. so I had to extract vdb data to .npy files.

1. Leaf Extraction
   Extract all non-empty 8x8x8 voxel blocks as well as the coords of each leaf origins (for reconstruction) from a dataset of VDB volumes.

2. VQ-VAE Model
   A PyTorch-based encoder compresses each leaf into a latent vector of size D in this case 128 dimmenions.
   The quantizer maps this vector to the closest of K learned codebook entries, the codebook as 256, which makes it fit in a uint8_t.

3. Loss Function
   Optimized using the following objective:

   $$\mathcal{L} = \|x - \hat{x}\|^2 + \beta \cdot \| \text{sg}[z_e(x)] - e \|^2$$

   with Exponential Moving Average (EMA) updates to the embedding table.

1. Load .vqvdb file, which contains:

   • Codebook (float32)
   • Per-leaf index tensors (4x4x4xuint8_t = 64 bytes)
   • Leaf origins (openvdb::Coord)

2. GPU Decoding

   • Launch CUDA kernel to decode latent codes into dense voxel blocks
   • Allocate leaf nodes in a new openvdb::FloatGrid
   • Write the reconstructed 8x8x8 voxel blocks

3. Streaming-Friendly

   • Decompression supports lazy loading in batches
   • Low VRAM footprint when streaming large scenes

• Hierarchical VQ-VAE (multi-res compression)
• Residual VAE
• Transformer-based latent upsampling
• Real-time decompression via Vulkan compute
• VDB segmentation / semantic-aware encoding

If you use VQVDB in academic work, please cite the project: