CoSTAR National Lab has open-sourced StarLightGS, an Unreal Engine 5.7 plugin that relights gaussian splats and casts their shadows without proxy meshes.
WallGS Pro brought splat relighting to Unreal Engine 5 in October, with shadow casting handled through an optional proxy mesh. StarLightGS instead renders relightable splats directly into the GBuffer, the intermediate stage that holds geometry and material data for Unreal's lighting and post processing passes. Relit splats receive every built-in light type, shadows from meshes and Lumen global illumination, so an emissive mesh will illuminate a nearby splat. They also contribute to screen space reflections, and relighting works with both Substrate and the legacy material pipeline.
The GBuffer stores only opaque surfaces, so StarLightGS renders relightable splats as opaque and offers an r.Gaussians.AlphaCutoff variable to trim the edge artifacts that can result. Splats do not feed the Lumen scene, so they cast no global illumination onto other objects. Ray tracing is unsupported, which rules out ray traced shadows, MegaLights and hardware Lumen.
StarLightGS shadows run through Virtual Shadow Maps. The plugin sends splat geometry to the shadow pipeline as a mesh batch and reshapes a quad for each splat, per light, so splats shadow meshes, other splats and themselves. Virtual Shadow Maps do not support translucent shadows, so splats cast as opaque ellipsoids with harder edges, and cost scales with the number of shadow casting lights.
Relighting requires PLY files with per splat normals, and metallic and roughness channels are optional. Box and sphere clipping volumes crop or segment a splat, and an Export Clipped Splat button saves the result as a new asset. PLY export writes either the original or Unreal coordinate space. The default ceiling is 5,000,000 splats, raised through r.Gaussians.MaxSplats as far as VRAM allows.
Relightable PLYs come from starlightgs-training, a Docker fork of GS-IR that runs COLMAP on photos or video and exports albedo, roughness, metallic and normals into a material PLY. The training pipeline needs an NVIDIA RTX 20 to 50 series, A100, H100 or H200 GPU on Linux or WSL2, and it is limited to noncommercial use because it bundles the Inria rasterizer. A sample project shows the plugin in a working level.
StarLightGS builds from source with Visual Studio 2022 or newer and is available now on GitHub as an MIT licensed beta.
