Abstract
With the advent of portable 360{\deg} cameras, panorama has gainedsignificant attention in applications like virtual reality (VR), virtual tours,robotics, and autonomous driving. As a result, wide-baseline panorama viewsynthesis has emerged as a vital task, where high resolution, fast inference,and memory efficiency are essential. Nevertheless, existing methods aretypically constrained to lower resolutions (512 $\times$ 1024) due to demandingmemory and computational requirements. In this paper, we present PanSplat, ageneralizable, feed-forward approach that efficiently supports resolution up to4K (2048 $\times$ 4096). Our approach features a tailored spherical 3D Gaussianpyramid with a Fibonacci lattice arrangement, enhancing image quality whilereducing information redundancy. To accommodate the demands of high resolution,we propose a pipeline that integrates a hierarchical spherical cost volume andGaussian heads with local operations, enabling two-step deferredbackpropagation for memory-efficient training on a single A100 GPU. Experimentsdemonstrate that PanSplat achieves state-of-the-art results with superiorefficiency and image quality across both synthetic and real-world datasets.Code will be available at \url{https://github.com/chengzhag/PanSplat}.