A new rendering framework separates the three things that make a 3D object look real — shape, light, and surface material — so each can be edited independently after the fact.
Existing Gaussian splatting methods bake lighting directly into the texture at capture time. That works fine for static scenes, but it falls apart the moment you try to move the object or change the light — the shading no longer makes physical sense. DR-GS (Deformable and Relightable Gaussian Splatting) fixes this by explicitly disentangling geometry, illumination, and material during reconstruction. The result is a pipeline where you can deform a captured object, swap the lighting environment, or tweak surface properties without re-running the whole capture process.
Gaussian splatting has been gaining ground in VR, AR, and digital content creation precisely because it renders fast and represents scenes explicitly rather than through a neural network's black box. The baked-lighting problem has been a persistent ceiling on its usefulness for production workflows, where artists need to composite assets into scenes with different lighting. DR-GS targets exactly that gap.
The paper reports strong results on glossy surfaces — specular highlights and reflections, which are notoriously hard to preserve under deformation — but peer review and independent testing will determine whether those gains hold outside the lab benchmarks it was evaluated on.