Image credit: 2025 Skylar Wurster
SIGGRAPH caught up with Skylar Wurster, a SIGGRAPH 2024 Posters contributor, to explore the groundbreaking research, “Gabor Splatting for High-Quality Gigapixel Image Representations.” This innovative technique leverages the periodic nature of Gabor functions to deliver exceptional parametric efficiency and reconstruction fidelity. Beyond image representation, Gabor splatting hints at exciting applications in video processing and scientific visualization. Here’s a closer look at the inspiration, advantages, and future potential of this cutting-edge work.
SIGGRAPH: Share an overview of “Gabor Splatting for High-Quality Gigapixel Image Representations.” What inspired this research? Can you elaborate on the periodic nature of Gabor splatting and how it contributes to improved parametric efficiency?
Skylar Wurster (SW): “Gabor Splatting for High-Quality Gigapixel Image Representations” is an approach that uses the periodic Gabor kernel to represent images at high fidelity while using fewer parameters than Gaussian splatting. We were inspired by the frequency domain’s success in representing signals, especially in popular compression techniques such as JPEG. We thought there was potential in the frequency domain for 2D and 3D representations but wanted to keep the spatial locality that Gaussian splatting uses for efficiency and explainability. Together, that leads us to Gabor kernels! The strength of the Gabor kernel is that with a single primitive, multiple gaussians in a periodic pattern can be represented, reducing the total parameter count for the representation with no degradation to quality.
SIGGRAPH: Beyond reconstruction quality, what other advantages does Gabor splatting offer over Gaussian splatting? Are there certain cases where Gaussian splatting or I-NGP might still outperform Gabor splatting, or is the latter universally superior for gigapixel image tasks?
SW: The reconstruction quality per parameter is our primary advantage. The compute is roughly the same, relying on somewhat similar CUDA kernels to splat the Gabor kernels. In our tests we found that our Gabor representation outperformed the other methods, but there may be certain scenarios untested that I-NGP or Gaussian splatting may outperform. For example, in very low parameter models, perhaps I-NGP will outperform Gabor splatting thanks to its efficient hierarchy of hash grids.
SIGGRAPH: Are there potential applications for Gabor splatting outside of image reconstruction, such as in video processing or scientific visualization? How do you see this work evolving over time?
SW: We believe a natural extension of this work is to radiance fields and 3D representation. We see our representation working in a tangential direction as other recent Gaussian splatting work, and there could be room for using Gabors as the kernel in current research directions for Gaussian splatting compression, 2D and textured Gaussian splatting, and Gaussian relighting. For scientific visualization, it could be useful to give another knob to turn besides the Gaussian position, scale, rotation, and color; now, periodicity can also be an added visual tool for detecting trends, outliers, and conveying a story with the data.
SIGGRAPH: What advice do you have for someone who wants to submit to Posters for a future SIGGRAPH conference?
SW: Don’t be afraid of failure, and be excited to share your work! Even if it is a smaller contribution, Posters is a great place to gain exposure and see where the idea can be taken next.
Feeling inspired? Share your innovations with the SIGGRAPH 2025 Posters program — submissions are open until Thursday, 24 April! Or, discover other exciting SIGGRAPH 2025 programs still accepting submissions and find the ideal fit for your work.
Skylar Wurster is a recent PhD graduate from The Ohio State University with a passion for 3D+AI and hopes to help democratize 3D content creation while also empowering professional 3D creators. His expertise spans 3D rendering, large-scale scientific data visualization, artificial intelligence, and radiance fields, with published papers in conferences such as SIGGRAPH and IEEE VIS on topics including Gaussian splatting, NeRF, hierarchical decomposition and super resolution, and fluid flow. Currently, he works on 3D generative AI on the Substance3D team at Adobe.
