GSOPs: 3DGS in Houdini

The integration of 3D Gaussian Splatting into additional platforms is progressing rapidly, with an already established presence in platforms like Unity and Unreal Engine. Today, we’re seeing SideFX’s Houdini receive access through GSOPs. GSOPs or Gaussian Splat Operators, comes from Creative Technologist, David Rhodes, and is completely free.

GSOPs allow for a range of operations, including isolating objects, noise and floater removal, model deformation and animation, scene composition, and generating meshes from gaussian splat models. It includes several SOP nodes, each designed to perform specific tasks like aligning, cropping, clustering, deforming, and exporting Gaussian splats. Additionally, users can perform statistical analyses of splat models and generate synthetic training data for Gaussian splat models.

GSOPs can create synthetic datasets for Gaussian Splatting

One of the big limitations in the research world I’ve come to learn is a lack of datasets for researchers to utilize. With GSOPs, this could be the beginning of the end of that problem. Using Houdini, people will be able to generate synthetic data with the hallmark view dependent features of radiance fields.

As part of the V1 release, here are the full list of available nodes:

1. gaussian_splats_align_by_points: This node aligns a gaussian splat model to the world origin, ensuring proper orientation and positioning for further editing or integration into larger scenes.

2. gaussian_splats_crop: Crops or groups a splat model.

3. gaussian_splats_dbscan: This tool performs density-based spatial clustering. It’s particularly useful for cleaning up models by removing noise and outliers, thus enhancing the clarity and quality of the final render.

4. gaussian_splats_deform: Deform splat models using polygonal geometry.

5. gaussian_splats_export: Exports Houdini gaussian splat geometry to disk, converting all relevant point data to native gaussian splat attributes in the process.

6. gaussian_splats_feature_analysis: Perform statistical analysis of gaussian splat models.

7. gaussian_splats_generate_training_data: Generate synthetic data suitable for training gaussian splat models.

8. gaussian_splats_import: Loads a trained gaussian splat model, converting all relevant data to native Houdini point attributes.

9. gaussian_splats_import_cameras: Imports the cameras.json file generated as a result of training gaussian splat models.

10. gaussian_splats_visualize_boxes: Visualize gaussian splats as opaque boxes.

GSOPs is pretty easy to install. Users need to clone the repository, copy the GSOPs.json file into the Houdini user preferences directory, and modify the file path accordingly. It comes with its own help card documentation and example scenes to guide users through gaussian splat workflows.

All digital assets and example scenes included on the install are under the Houdini Apprentice license, making them accessible for educational and experimental use. However, GSOPs itself is MIT Licensed for those that want to contribute, fork it, or explore it for their own purposes.

A viewport renderer is in progress, but in the interim, GSOPs will work with PlayCanvas’s Supersplat. There are also some known issues like rotation not updating spherical harmonics data and scaling issues leading to artifacts. However, Rhodes is working on these areas and encourages community contributions to enhance GSOPs further.

GSOPs has also been entered to SideFX’s H20 SIDEFX LABS Tech Art Challenge and it would mean a lot to David to vote for GSOPs, if you find it useful. For further inquiries or collaboration opportunities related to GSOPs, David Rhodes can be contacted at david.a.rhodes.3d@gmail.com or through his Linkedin.

Another major platform now supports a radiance field method. What other applications are on your bucket list?