DENS-E fits 3D Gaussians to OpenVDB volumes — smoke, clouds, fire, tornadoes, and arbitrary density grids — with every step of training running on the GPU. Forward and backward passes, the optimisers, adaptive density control, metrics, and both screen-space splatting and 3D Gaussian-to-volume rasterization are implemented as Slang compute shaders dispatched through slangpy on Vulkan. The output is a compact set of anisotropic Gaussians that reconstructs the source volume, renders in real time, and slots into existing Gaussian-splatting tooling.
All the results of the experiments presented in the diploma are here:
More information regarding the project can be found on Github:
https://github.com/Adeon18/DENS-E
Reference vs reconstruction
The application feature showcase is presented in this video along with a training process:
https://drive.google.com/file/d/1THwDXA_thcifXJETWWx5d5Zupa52fe3T/view?usp=sharing