Stefan VaskevichSix months ago I was under the impression that 3D Gaussian Splats were a dead end. Cool looking, but useless for real games. No collision, no level of detail, no engine plugin worth using. Then two demos landed in the same week, and I had to change my mind.
First, a fully playable FPS in the browser where the whole environment is a single photogrammetric splat, with collision, AI enemies, and shooting. Second, an open-source toolkit that takes one image and gives you a 3D world you can walk inside in five minutes, including objects you can throw around that actually collide with the geometry. Something flipped. Here is the full state of playable Gaussian Splats in 2026: the generators, the engine plugins, the collider trick, and the hardware you actually need.
Big credit on this one to Gabriele Romagnoli, a tech ambassador deep in the 3DGS space who helped me map the landscape.
Key Takeaways
- ▸Four 2026 unlocks closed the gap: mesh colliders shipping with splats (KIRI 4.2, Marble), NanoGS Nanite-style LOD for Unreal, Niantic SPZ 4 files 10× smaller, and the glTF KHR_gaussian_splatting standard.
- ▸Marble (World Labs) leads world generation with a dedicated collider GLB export; Hunyuan World 2.0 is the free open-source rival for offline workflows.
- ▸One winner per engine: NanoGS in Unreal, UnityGaussianSplatting (aras-p) in Unity, Spark.js in Three.js + WebXR. The web stack was the most popular at the SensAI Hackathon (54% of teams).
- ▸Free starter stack: Scaniverse to capture → SuperSplat to clean →
splat-transform --collision-meshfor the collider → NanoGS or Spark.js to render. Zero dollars, one afternoon.
What changed in 2026? The four unlocks
A Gaussian Splat is a cloud of millions of tiny ellipsoids: each with a position, a color, an opacity, and a shape. Not triangles. Alpha-blended blobs that the GPU rasterizes from any angle. Beautiful for backgrounds, useless for gameplay, because you cannot walk into a cloud of color. Until this year.
Four things shipped in the last six months that, together, closed the gap:
1. Mesh colliders shipped with the splat
KIRI Engine 4.2 (April 2026) and World Labs Marble now export a dedicated collider GLB next to the splat. Invisible mesh handles physics, the splat handles visuals.
2. Nanite-style LOD in Unreal
NanoGS (March 2026, MIT, UE 5.6+) is the first Nanite-class level-of-detail plugin for splats. ~4× framerate uplift on an RTX 2070 (per the NanoGS README benchmarks, accessed May 2026).
3. Files got 10× smaller
Niantic's SPZ 4 format (May 2026) compresses an apartment-scale scan from a ~500 MB raw .ply down to 50–80 MB. Loads in roughly three seconds.
4. The glTF standard caught up
The KHR_gaussian_splatting extension hit Release Candidate in February 2026, with NVIDIA, Google, Adobe, and Cesium aligned. Splats are now a first-class citizen of the 3D web.
The catchy demo: image-blaster
The post that put splats back on my radar came from @theworldlabs on 14 May 2026, with 98K views and 600 bookmarks in two days. Title: “turn a single image into a fully meshed 3D world in minutes.” The project is called image-blaster and it's sitting on 2.9K stars at the time of writing. The author, Neilson Koerner-Safrata, is on the World Labs design team, so this is essentially a first-party reference implementation, not a hack.
The stack is a Claude skill that orchestrates three APIs in one shot: Marble 1.1 for the environment splat, Hunyuan-3D via FAL for props (meshed, not splatted, that's the trick), and ElevenLabs for ambient SFX. You drop an image into /input/, tell Claude “blast it,” and five minutes later you have an .spz environment, a folder of .glb meshes, and an .mp3 of room tone. Drop the bundle into Unity, Unreal, Godot, Blender, or Three.js.
Honest report after running it on my own footage: install was almost instant. You only need to set the World Labs and FAL API keys in the .env file and ask Claude to install. One environment generation with a handful of objects costs roughly a dollar in credits. The output viewer lets you walk around immediately, and the spawned objects do collide with each other. The environment collision still needs manual work, which is the next thing we'll fix.
Why this is suddenly possible: Marble's collider export
The engine underneath image-blaster is Marble, World Labs' first commercial product. It's Fei-Fei Li's company, and they closed a $1B round in February 2026 (TechCrunch, accessed May 2026), with a $200M check from Autodesk alongside backers like AMD, NVIDIA, and Fidelity. So the “splat” tooling is now bankrolled by the company that builds Maya, 3ds Max, and Revit, plus the people making the GPUs that render it.
Marble 1.1 (default since April 2) ships export guides for Unreal, Unity, Houdini, Blender, and Three.js, including a WebXR preset for Quest 3 and Vision Pro. The killer feature for game work is the dedicated collider GLB it exports next to the splat: an invisible watertight mesh you can drop into your physics system while the splat renders on top.
Which world generator should you pick?
Marble is not alone. Here are the five I would actually consider this month, with what each one is good at:
| Tool | Type | Why it's on the list |
|---|---|---|
| Marble | Commercial | The leader. Only generator with a dedicated collider GLB export. WebXR-ready. Dominated the SensAI Hackathon (94% of world-model teams). |
| Hunyuan World 2.0 | Open source | Tencent, open-sourced 16 April 2026. Outputs mesh + splat + point cloud. Direct UE/Unity/Isaac export, offline. The free local rival. |
| Genie 3 | Google DeepMind | Real-time playable worlds, but no export and far less controllable. AI Ultra tier, $250/mo. Hype was loud, use cases have been quiet. |
| Odyssey-2 Max | Commercial | Native splat output at ~20 fps. Pixar's Ed Catmull is on the board. Strong for real-time creative workflows. |
| Blockade Labs Skybox | Free | The OG 360° skybox tool. Free tier. World Mesh GLB export in alpha. The instant low-fi fallback. |
Worth calling out: Hunyuan World 2.0 has built-in TSDF + marching-cubes mesh extraction, so collision works offline, out of the box, on your own GPU. If you cannot pay for Marble credits, this is the drop-in replacement. (Tencent's broader Hunyuan family has been on a tear this year. See our Hunyuan 3D tool page for the object-generation side.)
The per-engine toolkit
One winner per platform, no plugin landscape salad. These are the ones actually used in production right now:
Unreal Engine 5 → NanoGS
Free, MIT, UE 5.6+. Nanite-style LOD. Shipped 19 March 2026. For animated 4DGS, pair with Dazai's SplatRenderer.
Unity → UnityGaussianSplatting
aras-p's reference implementation. MIT. The default the SensAI mentors recommended for multi-user and physics work.
Three.js + WebXR → Spark.js
Built by World Labs. Supports .ply, .spz, .splat, .ksplat, and .sog. Handles WebXR on Quest 3 and Vision Pro out of the box. 54% of SensAI Hackathon teams used it (March 2026 event report), more than any other integration path.
The collider trick (works on all three)
Pair the splat with a KIRI Mesh-Inclusive capture or a Marble collider GLB. Set the mesh invisible, keep collision and navmesh on it. The splat is just paint.
From splat to videogame: the PlayCanvas FPS
The cleanest end-to-end proof is PlayCanvas's “Turning a Gaussian Splat into a Videogame” (PlayCanvas blog, 22 April 2026, accessed May 2026). They took a real abandoned location scanned by Christoph Schindelar and built a first-person shooter on top: eight AI-driven NPCs, weapon fire, and a navmesh, all running in your browser, free.
The trick is a CLI tool called splat-transform with a --collision-mesh flag. It voxelizes the splat, flood-fills the empty volume, and exports a watertight .collision.glb. Feed that to Recast for the navmesh and you have a playable level.
Capture your own: scan-to-splat tools
If you want to bring real spaces into your game instead of generating them, the three I would recommend by default:
| Tool | Cost | Why pick it |
|---|---|---|
| Scaniverse | Free | Niantic. Fully on-device, iOS + Android. Cheapest path to a usable splat: zero subscription, no upload required. |
| KIRI Engine 4.2 | Free tier | 3DGS-to-Mesh 3.0 (April 2026): one capture gives you the .ply splat AND an .obj/.fbx collision-ready mesh in the same coordinate space. |
| Polycam | $8/mo Pro | iPhone LiDAR support. Best quality-to-effort if you have the hardware. |
Desktop bonus: Postshot is free (post-beta, Windows + NVIDIA). Pure local OSS: gsplat via Nerfstudio runs on a 4 GB+ GPU. Heads-up: Luma iOS dropped direct 3DGS support in 2026, so don't recommend it for new captures anymore.
If you want the single-image, generate-from-photo path instead of a phone scan, our walk-through of the EasyEnv Blender add-on covers the same idea on a smaller scale, see Image to 3D World in 60 Seconds.
To browse existing scans rather than make your own, three libraries worth knowing: Into the Scaniverse (50K+ user splats on Quest + WebXR, free), Polycam Explore, and the in-browser editor SuperSplat for cleaning, cropping, and optimizing whatever you grab.
What hardware do you need to run Gaussian Splats?
The viewing side is the surprise. The entry bar is lower than most people assume:
- Any modern phone runs SuperSplat in the browser.
- A mid-range desktop GPU (RTX 3060) holds roughly 60 fps on <1M Gaussiansin my testing with apartment-scale captures at SPZ 4 compression.
- One million Gaussians takes roughly 180–220 MB of VRAM and 3–8 ms per frame on an RTX 3070, in line with the numbers published in the gsplat / Nerfstudio benchmarks.
Training your own splats is the other side. gsplat and Brush (Niantic, OSS) both start at 4 GB VRAM. Brush is the interesting one: it runs in a browser via WebGPU. No install, no CUDA, no Python environment.
Cheapest end-to-end path with zero specialty hardware: Scaniverse on your phone → SPZ 4 export → SuperSplat to clean → Spark.js in a browser. No GPU, no install, no subscription.
The verdict
The honest summary: photoreal capture is no longer a parlor trick for VFX backgrounds. With the collider GLB workflow plus NanoGS or UnityGaussianSplatting, you can ship a level built from a splat, today. The PlayCanvas FPS proves it.
Where I'd still hesitate: anywhere you need dynamic global illumination, animated characters embedded in the scene, or competitive-shooter precision. Splats render as alpha, not a full PBR pass, and that's a real trade-off that won't close this year.
How this was tested. All hands-on claims in this article (image-blaster install, Marble generation cost, walkable output, hardware framerates) are from my own tests in May 2026 on a Windows desktop with an RTX 3070 and a consumer iPhone capture pipeline. Third-party numbers (Marble pricing, NanoGS uplift, SensAI Hackathon stats, SPZ 4 compression) are linked to their primary sources with an access date. TOP 3D AI does not take payment for rankings or coverage; see our editorial independence policy.
Frequently asked questions
Can you use 3D Gaussian Splats inside Unreal Engine or Unity?
Yes, as of 2026. NanoGS (free, MIT, UE 5.6+) ships Nanite-style level-of-detail for splats inside Unreal. UnityGaussianSplatting by aras-p is the reference plugin for Unity. Both render the splat visually; you pair it with a separate invisible collider mesh (KIRI Mesh-Inclusive or Marble's collider GLB) for physics and navmesh.
How do you add collision to a Gaussian Splat?
Three ways. (1) Generate the world with Marble, which exports a dedicated collider GLB alongside the splat. (2) Scan with KIRI Engine 4.2, whose 3DGS-to-Mesh 3.0 outputs the .ply splat and an .obj/.fbx collision mesh in the same coordinate space. (3) Use PlayCanvas's splat-transform CLI with the --collision-mesh flag, which voxelizes the splat, flood-fills the empty volume, and exports a watertight .collision.glb. In all three, the mesh is set invisible and the splat renders on top.
How large is a typical Gaussian Splat file in 2026?
Niantic's SPZ 4 format (May 2026) compresses an apartment-scale scan from roughly 500 MB raw .ply down to 50–80 MB. Compression is 3–5× faster than older formats, and the 10-million-point ceiling is gone.
What hardware do I need to view or train Gaussian Splats?
Viewing is cheap. Any modern phone runs SuperSplat in the browser. A mid-range desktop GPU (RTX 3060) holds 60 fps on under 1M Gaussians. Training your own splats starts at 4 GB VRAM with gsplat (Nerfstudio) or Brush (Niantic, OSS). Brush also runs in the browser via WebGPU with no install.
Marble vs Hunyuan World 2.0: which world generator should I pick?
Marble (commercial) is the production leader because it is the only generator that exports a dedicated collider GLB and ships first-party export guides for Unreal, Unity, Houdini, Blender, and Three.js. It dominated the SensAI Hackathon at 94% of world-model teams. Hunyuan World 2.0 is the free local alternative with built-in TSDF + marching-cubes mesh extraction, so collision works offline on your own GPU.
What is the cheapest end-to-end path to ship a playable splat scene?
Scaniverse (free, on-device, iOS or Android) to capture, SPZ 4 export, SuperSplat to clean, then splat-transform --collision-mesh for the collider, and finally NanoGS in Unreal or Spark.js in the browser to render. Zero dollars, one afternoon.
Want to track who is actually winning at AI 3D generation right now?
Compare AI 3D tools in the ArenaOr see the current ELO leaderboard for blind-vote rankings across every major model.
