Gaussian Splats Cast Shadows Now

Shadows were the tell. You could always spot a Gaussian Splatting scene in a demo reel because everything floated in ambient light — beautiful geometry, gorgeous color, zero ground contact. NVIDIA's vkSplatting 2026.1, dropped at GTC and now on GitHub under Apache 2.0, fixes that. And honestly, the shadow support is the least interesting part of the release.

#What vkSplatting 2026.1 Actually Ships

Three things matter in this update.

Ray-traced shadows with DLSS Ray Reconstruction. Gaussian Splatting doesn't naturally encode light direction — each splat is a colored ellipsoid with opacity, not a surface with a normal. vkSplatting works around this by casting rays from a directional or point light source and testing intersections against the splat field. DLSS Ray Reconstruction handles the denoising, which is necessary because tracing rays through millions of semi-transparent blobs produces noisy results at interactive rates. The shadows aren't physically perfect — they're approximations — but they're convincing enough that a scanned environment no longer looks pasted onto a blank void.

Multi-instance splat set architecture. This is the feature that actually matters for production. You define a splat asset once and instance it across a scene using global index tables and unified sorting. Think of a city block where the same window frame appears 500 times. Before this release, that meant 500 copies of those Gaussians eating VRAM. Now it's one copy with 500 transforms. Memory consumption drops by an order of magnitude for repetitive environments — which describes most built environments humans actually inhabit.

Centralized bindless asset system. Less exciting to explain, more exciting to use. Scene load times and memory allocation become predictable. For anyone building a pipeline around splats rather than just viewing them in a one-off demo, predictability matters more than peak throughput.

The renderer is Vulkan-native, open source, and lives on NVIDIA's GitHub alongside their other DesignWorks samples. You need an RTX card for the ray-traced shadows and DLSS path. AMD users still get instancing and the bindless system but lose the shadow quality.

#The Engine Situation

The question everyone asks after a renderer demo: "cool, but can I ship this in Unreal?" As of March 2026, the answer is a qualified yes — and the qualifications are shrinking fast.

NanoGS deserves its own paragraph. It borrows the core idea behind Nanite — automatic LOD clustering — and applies it to Gaussian splat fields. A scene with 10 million Gaussians gets broken into hierarchical clusters, and the renderer streams only the detail level the camera actually needs. Early tests show 60 fps on scenes that previously required aggressive manual decimation to run at all. If NanoGS matures through the rest of 2026, it removes the single biggest practical obstacle to shipping splat-heavy content in games: the "it runs great on my 4090, good luck on anything else" problem.

The Cesium integration is the sleeper development. KHR_gaussian_splatting as a glTF extension means splats are no longer trapped in PLY and SPLAT files. They travel through the same asset pipelines as meshes, materials, and animations. Digital twins, geographic information systems, architectural visualization — these industries were waiting for exactly this kind of standardization before writing purchase orders. A glTF file that contains both traditional geometry and Gaussian splat data, loadable in a browser or a game engine with the same API call? That's infrastructure, not a demo.

#Where the Gaps Still Are

Not everything is solved, and some of the remaining problems are structural.

Real-time editing of splat scenes is still painful. You can't grab a cluster of Gaussians and reposition them the way you'd move a vertex group. The representation is statistical, not geometric, and direct manipulation feels like trying to sculpt fog. Tools like SuperSplat offer basic cleanup — deleting floaters, cropping bounds — but anything resembling creative editing requires re-training the scene from modified source images.

Material reassignment is the other hard wall. A Gaussian splat bakes appearance at scan time. Want to change the wall color from beige to blue? You're either retraining from scratch or layering a post-process hack that never quite looks right at grazing angles. Mesh-based workflows still dominate on art direction flexibility, and that gap won't close until someone figures out how to decompose splat appearance into something like a PBR material model.

VRAM remains the constraint nobody at keynotes wants to dwell on. A detailed room scan at full fidelity can hit 4–6 GB. A city block pushes past 12 GB before instancing, and instancing only helps when geometry repeats. Organic environments — forests, rocky coastlines, cluttered studio apartments — don't instance well.

#The Practical Read

Here's where I'd reach for vkSplatting 2026.1 today: scanned built environments for real-time architectural walkthroughs or virtual production scouting. The instancing handles repeated structural elements — windows, floor tiles, light fixtures. The shadows ground the scene on its surfaces. The Vulkan renderer gives enough performance headroom to stream on a workstation GPU without precomputing everything into baked lightmaps.

Where I wouldn't use it yet: anything requiring post-scan art direction, game levels with dynamic lighting that shifts per scene, or mobile delivery targets. The stack is desktop-first and capture-first. If your workflow starts from concept art rather than a LiDAR scan, you still want meshes.

The Superman production reportedly used dynamic Gaussian Splatting for set extensions — first major film to ship that way. That isn't vkSplatting specifically, but it signals where industry gravity is pulling. When VFX houses start baking splats into their dailies pipelines, tooling follows money. Give it another year and "splats can't handle lighting" will sound as quaint as "GPUs can't do ray tracing" did five years ago.