Reconstruction par étalage gaussien 3D de phénomènes gazeux pour les effets visuels

Abstract

Rendering volumetric phenomena such as smoke, fire, vapour, and clouds require prohibitive computation times. Meanwhile, 3D gaussian splatting-based (3DGS) methods are an emergent family of differenciable rendering algorithms which allow to capture and subsequently render a subject’s radiance from novel viewpoints at unmatched interactive framerates and quality. However, only a few projects researched the possibility of leveraging 3DGS in the context of volumetric visual effects (VFX) productions. We think that 3DGS can be used to transcode (also known as baking) volumes to a 3DGS representation to accelerate post-authoring renders. The present master’s thesis proposes such a method. Our method allows to circumscribe the computationally prohibitive light transport simulation to the volumetric VFX authoring step. We achieve this by proposing modifications to the 3DGS method in the initialization and optimization steps. We improve the initialization by leveraging properties of standard volumetric storage formats and of the camera parametrization used in VFX authoring software; this addresses the issues that arise when using photogrammetric registration, as proposed by the 3DGS reference implementation, in our volumetric context. To specifically improve the reconstruction of low opacity volumetric subjects, we propose anisotropy and position regularizations, an implementation of the quadratic splatting kernel as an alternative to the gaussian kernel, and a multi-background rendering and evalution algorithm. We test the conclusions of other gaussian reconstruction, modelization, and rendering research endeavours when applied in our context. Our results show that our method allows the reconstruction of volumetric subjects with greater photometric accuracy than existing gaussian reconstruction methods. Our method produces gaussian models which are more geometrically cohesive and are better separated from their training environment. Consequently, the models are usable in a compositing workflow. Our gaussian models’ transparency correctly encode their respective ground-truth’s transparency, allowing them to be composited with other gaussian models, with polygonal models, and with environment maps featuring similar lighting conditions as the ground truth, as demonstrated by our results.

Date	9 Jun 2026
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Eric Paquette (Supervisor)