Still2Scene: Hybrid Gaussian Environments for Virtual Production

Abstract

Virtual production often requires rapidly generated background environments that support real-time rendering and limited camera motion. While volumetric 3D Gaussian splatting provides high visual fidelity, it is computationally expensive, whereas planar billboard representations are efficient but lack geometric depth. We present a hybrid Gaussian scene representation that converts a single image into a lightweight navigable environment for virtual production. The proposed approach combines volumetric foreground Gaussians with planar Gaussian primitives for distant regions, while additional billboard assets can be synthesized using diffusion-based image prompting. The resulting hybrid scene can be deployed directly in Unreal® Engine for real-time rendering. This workflow enables fast environment generation from images and provides a practical middle-ground between static backplates and fully authored 3D environments in virtual production pipelines.