Relighting Gaussian Splats is Hard. Here's why it's one of the biggest unsolved problems in 3DGS, and how one team is tackling it. When you capture a Gaussian Splat, you capture the light of the scene baked right in: the shadows, the highlights, the color temperature. All of it is "locked" inside the splat. This is what Fernando Rivas-Manzaneque at Volinga calls the additive lighting problem. In the interview at XR AI Spotlight he told me: "You are always adding light on top of what was captured. The original lighting stays. You layer on top of it" Two approaches are being explored to get around this: - Pre-delighting the source material. Teams like Beeble AI strip harsh shadows from the raw footage before the splat is ever trained. Clean input means a clean base to relight from scratch. - Mesh-based lighting. A proxy mesh is attached to sections of the splat, giving it real material properties and normals. Lights in your scene can then interact with it physically. Both require extra work. One before training. One inside the engine. Neither is a one-click solution yet. But as Gaussian Splatting moves deeper into virtual production and episodic VFX (VOLINGA recently helped CBS/Paramount+ rebuild a silo environment for Fire Country without a single dangerous scaffold), relighting becomes the feature needed to unlock the whole format for professional use. Make sure to check out the full interview on my YouTube Channel