Closed-Form Convolution for physically-accurate defocus in Gaussian Splatting

Closed-form convolution is being used to model physically accurate defocus in Gaussian splatting, pushing the technique beyond the usual approximation-heavy treatment of depth of field. For developers working with splats, the practical appeal is obvious: defocus can be integrated into the representation itself instead of layered on afterward as a separate blur effect.

That matters because Gaussian splatting has become attractive for fast scene reconstruction and rendering, but optical effects like depth of field are often where approximation quality starts to show. A closed-form approach can reduce artifacts, preserve shape and energy better across focus changes, and make the result more predictable for real-time or near-real-time use.

The work appears in ACM Transactions on Graphics, Volume 46, Issue 1, with publication dated February 2027. The exact implementation details and performance tradeoffs will determine how practical this is for production pipelines, but the direction is important for anyone building splat-based renderers, capture tools, or hybrid image-based workflows.

For game teams, the bigger question is whether this can be folded into existing renderers without blowing up cost or complexity. If it can, physically grounded defocus could become one of the more visible quality wins for splat-driven scenes, especially in cinematic cameras and reconstruction-heavy content.