Did you know that Gaussian splatting always assumed that none of the Gaussians overlap, which is definitely not a correct assumption? Our CVPR 2026 (findings) effort Softmax-GS improves over the non-overlapping assumption in Gaussian splatting, and make multiple Gaussians compete in regions where they overlap using a softmax function, which greatly improves rendering performance when less Gaussians are used, as well as rendering consistency from different viewpoints. Especially, check out our experiments fitting multiple geometric shapes with only 4 Gaussians!
https://arthurhero.github.io/projects/smgs/