Gaussian Splatting: Community News #1
Stay informed about all the latest developments in the Gaussian Splatting ecosystem🤘
Luma AI now available on Android
Luma AI has now expanded its offerings by launching its app for Android users. Experience the enhanced features and user-friendly interface on your Android device today: https://play.google.com/store/apps/details?id=ai.lumalabs.polar
GSOP Update
Interested in using Gaussian Splatting for motion graphics, animation, or VFX projects? David Rhodes has added instructions for working with splat (animation) sequences in GSOPs and Unity: https://github.com/david-rhodes/GSOPs/blob/develop/extra/UnityGaussianSplatting/INSTRUCTIONS.md
This is a simple example, but the potential is endless. Because Splats are represented as points in Houdini, they are very easy to manipulate. As a reminder, you can also rig and articulate Splat models with GSOPs. Check David’s post on LinkedIn: https://lnkd.in/gZGR3qgX
You can download a commercial version of GSOPs here: https://lnkd.in/gzH-XYtc (starts download immediately)
Reducing the File Size of Gaussian Splats
The file size associated with Gaussian Splatting is becoming increasingly crucial for several reasons. Firstly, the scale of 3D models generated using Gaussian Splatting is expanding, with the technology now capable of processing entire cities in one go, leading to substantially larger datasets. Secondly, the advent of 4D Gaussian Splatting (experience here:https://antimatter15.com/splaTV)/, which extends the technology to 3D video content, further multiplies the data involved, necessitating more efficient handling and storage solutions to manage these enhanced file sizes.
The original authors of the 3DGS paper have returned with an approach to significantly reduce the file size of Gaussian Splatting.
Project website: https://repo-sam.inria.fr/fungraph/reduced_3dgs/
Background information: https://radiancefields.com/shrinking-3dgs-file-size
GauStudio: A framework dedicated to Gaussian Splatting
GauStudio is a joint research effort of the Shenzhen Institutes of Advanced Technology, Chinese University of Hong Kong, Shenzhen, and Technical University of Munich. Here is why we think this framework dedicated to Gaussian Splatting is very promising:
- They are the first to unveil the technical details behind the ‘skyball’ technique employed in Luma AI’s 3DGS algorithm. This technique is instrumental in separating foreground and background elements in outdoor scenes. Their report and the upcoming code release will enable many companies to achieve similar results to Luma AI when processing outdoor imagery.
- They have reimplemented the appearance embedding technique from VastGaussian and combined it with their Gaussian Surface Reconstruction (GauS) technology in GauStudio. This has enables Gaussian Splatting, for the first time, to reconstruct a mesh from a very large outdoor scene. No other NeRF or Neus-based method has been capable of achieving such fine detailed shapes for large-scale scenes, comparable to what multi-view stereo (MVS) methods can produce.
- Since GauStudio supports various 3DGS-based methods like Scaffold-GS, GauHuman, and even 2DGS in the near future, the GauS method in GauStudio is the only toolkit that can be applied to every GS-based method to recover surfaces.
- Given GauStudio’s modular architecture, it is straightforward to replace components like the renderer from the original 3DGS rasterizer with commercial-friendly alternatives such as gsplat or OpenSplat.
- They plan to offer commercial-free substitute modules for companies, ensuring that these entities can utilize GauStudio without any licensing conflicts.
Learn more about GauStudio on GitHub: https://github.com/GAP-LAB-CUHK-SZ/gaustudio
