Spotlight: Unconstrained Text-to-3D Scene Generation with Panoramic Gaussian Splatting

In a recently published paper, UCLA undergraduate ECE student Tejas Bharadwaj reflects on being a co-author of a top-tier CS conference paper. Tejas is a co-inventor of DreamScene360 a technique which explores “…a text-to-3D 360 scene generation pipeline that facilitates the creation of comprehensive 360 scenes for in-the-wild environments in a matter of minutes.” Using the generative power of a 2D diffusion model and prompt self-refinement, the approach creates a high-quality and globally coherent panoramic image that acts as a preliminary “flat” (2D) scene representation. The 2D scene representation is “…lifted into 3D Gaussians, employing splatting techniques to enable real-time exploration.” 

Reflecting on his experience in publishing, “Computer vision has always interested me since I had an intuition for math and geometry, ” Bharadwaj explains. “… So I did an internship at Prof. Kadambi’s lab over the summer of my freshman year. After I got up to speed on the basics, I wanted to work on something novel and gen-AI was and still is a booming field so my PhD mentor and I took part in brainstorming ideas for a paper, although my contribution was mostly on the implementation side. The paper is another small step in the massive AI boom and I hope others and ourselves can find new better algorithms for 3D gen-AI, and from a personal standpoint, I learned quite a bit from contributing to this research both about the field of vision/AI as well as the overall scientific process.” 

Bharadwaj’s method offers a “globally consistent 3D scene within a 360 perspective, providing an enhanced immersive experience over existing techniques.”

Read more about Tejas Bharadwaj’s research here: https://dreamscene360.github.io/