 |
Abstract: Integration of Computer-Generated Content in Gaussian Splatting Scenes: Shadow Catching and Hybrid Light Estimation [DE]
Tobias Preuss, Integration of Computer-Generated Content in Gaussian Splatting Scenes: Shadow Catching and Hybrid Light Estimation, Hochschule Düsseldorf - University of Applied Sciences, Master thesis, 05.05.2026.
Gaussian Splatting is revolutionizing 3D reconstruction through efficient real-time
rasterization. For physically accurate image synthesis, however, the visual
effects industry relies
on polygon-based ray tracing. This fundamental incompatibility has so far prevented the
seamless integration of Gaussian Splatting scenes into established production pipelines.
This thesis presents a hybrid approach for evaluating and integrating computer-generated
objects into splat environments. Consciously avoiding the development of proprietary render
engines, the application-oriented workflow exclusively utilizes established commercial tools
(2DGS, Maxon Redshift and Foundry Nuke). The technical implementation of the workflow
is based on three core components: first, utilizing the splat as a spatial, emissive light source
for near-field radiometric information; second, supplementing it with traditional CG lights
for targeted main illumination; and third, applying differential rendering for the correct
extraction and final composition of the render passes.
The results, validated through expert interviews, demonstrate strengths in depth-based light
interactions: spatial parallax and spherical harmonics considerably enhance the realism of
reflections. Furthermore, the method proves to be highly effective in organic environments
and unrestricted camera framing. Nevertheless, clear limitations emerge: the lack of closed
geometry reduces shadow precision, while the absence of native BRDF data and the 8-bit
limitation restrict exact light interactions. Moreover, the workflow inevitably reproduces
existing splatting artifacts, making the quality of the integration highly dependent on the
inherent characteristics of the splat environment.
In conclusion, the hybrid workflow offers immense spatial flexibility but requires a
scenario-specific evaluation against the precision of traditional projection methods.
This thesis demonstrates the methodological feasibility and paves the way for future GS implementations
in post-production.
Keywords:
Gaussian Splatting, Visual Effects, CGI Integration, Hybrid Rendering, Differential Rendering, Ray Tracing
Supervisor:
Prof. Jens Herder, Dr. Eng./Univ. of Tsukuba
Prof. Dr.-Ing. Sina Mostafawy
Location:
The research took place at the Virtual Sets and Virtual Environments Laboratory.
Video
|
|
