Hey everyone,
If you've ever battled with multi-gigabyte VDB caches from pyro, I've been working on a project you might find interesting.
I'd like to introduce VQVDB, a free and open-source plugin for Houdini that uses a VQ-VAE neural network to compress OpenVDB float grids. The goal is to dramatically reduce the file size, up to 32x, of simulation caches, making them easier to store and faster to transfer.
You can grab it right now on GitHub: https://github.com/ZephirFXEC/VQVDB/releases [github.com]
The technical paper is available : https://doi.org/10.6084/m9.figshare.29469083.v3 [doi.org]
Here’s a quick comparison showing the visual quality next to the original. The compression can achieve significant file size reduction while maintaining structural and visual detail.
And here’s a look at the simple workflow using the provided SOPs nodes.
This is an early release with some limitations:
- OS: Windows only (for now).
- Hardware: An NVIDIA GPU with CUDA support is strongly recommended, CPU is untested.
- Grid Type: Currently supports FloatGrid (scalar float) only.
This is just the beginning!
I'm planning to explore:
- Linux support.
- Remove dependencies and write GPU Agnostic backend
- Support for vector grids (Vec3fGrid).
This is an early release, and I'd love to get your feedback, hear your ideas, and find out about any bugs you encounter.
Feel free to open an Issue on the GitHub page.
Hope you find it useful!
Enzo C.
