Is Gaussian Splatting the game-changer our 3D models have been craving?
We've been digging into Gaussian Splatting having seen impressive results on YouTube and across our social media platforms. We wanted to investigate what quality of 3D output was available from the Gaussian Splatting technique and how it compares to some of our existing photogrammetry workflows in terms of speed and quality.
Initial tests.
We kicked things off by filming a 360° video of one of our Herman Miller office chairs from different angles. First, we used a quick 20-second clip from a phone camera and processed this using Luma AI's free web tool. We were pleasantly surprised by the results, especially considering that we didn't use any specialist lighting or setups and it took seconds to capture.
Increasing quality.
Intrigued, we decided to step it up with a second video, using a fancier 4K camera and a much longer video recording. The difference was clear – the details were much sharper, especially on the chair's back assembly. However, we also encountered a quality issue, as the improved detail in the back assembly was overshadowed by the terrible geometry of the mesh seat area. It seems that the software struggled to handle the transparency of the fine mesh effectively.
Is Gaussian Splatting the future of 3D capture?
We were pleasantly surprised by the results, especially considering that we didn't use any specialist lighting or setups. The Luma AI site provides .glb and .obj files, which are perfect for importing for use in other 3D software. And while it might not match the quality of traditional 3D scans, it's great for creating a base mesh, especially soft bodies like fabrics and cushions, where our 3D Artists can take time to refine the model. Plus, we're thinking it could be handy for clients – they could send us a quick video and we could turn it into a model the very same day to demonstrate the benefits of CGI on their specific project.
So, Gaussian Splatting might just be the boost we need for some of our 3D projects.