// 2026.05.22

From an Image to a Headset Scene: image-blaster → Unity/Unreal → Quest 3 & Vision Pro

10:00 JST

Need an environment fast? image-blaster (MIT) is a Claude Code skillset that turns a single image into a meshed 3D world in under five minutes. This post covers the engine route to a headset — the heavier path, but the one with the best on-device performance and full XR input. (For the lighter browser route, see our A-Frame / WebXR starter.)

What image-blaster gives you

Run claude in the repo, supply World Labs and FAL API keys, drop an image in input/, and ask Claude to “blast it and confirm each step with me.” You get:

  1. .glb / .obj meshes of the dynamic objects (Hunyuan-3D via FAL — tune --face-count 40000-1500000, default 50000; --generate-type Normal|LowPoly|Geometry)
  2. .spz Gaussian splat of the static environment (World Labs Marble marble-1.1)
  3. .mp3 ambient loop + per-object SFX (ElevenLabs)

The skillset is MIT, but it calls paid third-party APIs (World Labs, FAL/Hunyuan, ElevenLabs) — bring API keys and credits. This is also why it pairs well with our 3DGS-from-360°-camera workflow: both end in a Gaussian splat you have to get on-device.

The engine pipeline (Path B)

1 · Generate. IMAGE-BLAST your reference image. For headsets, generate meshes with a sane budget — keep --face-count modest (≈50k or lower) and consider LowPoly for dynamic props.

2 · Import meshes. .glb/.obj import natively into Unity, Unreal, or Godot. These are your interactive/dynamic objects and colliders.

3 · Handle the splat. The .spz environment is the hard part — engines don’t render Gaussian splats natively. Convert .spz → .ply (SuperSplat / spz tooling), then add a splat renderer:

  • Unity → Quest 3: use a Gaussian-splatting package (e.g. Aras Pranckevičius’ UnityGaussianSplatting). It’s compute-shader based and mobile/Quest support is experimental — keep splat counts low, test early, and have a mesh fallback for the environment.
  • Unreal → Quest 3: a Gaussian-splat plugin + OpenXR; same perf discipline.
  • Apple Vision Pro: the toughest target — RealityKit has no built-in Gaussian-splat rendering, and Unity PolySpatial/visionOS doesn’t render splats out of the box. Pragmatic options: (a) ask World Labs Marble for a mesh export of the environment and skip splats entirely on AVP, or (b) use a Metal-based splat renderer in a full-immersive Unity/visionOS app. The mesh-only route is by far the faster hackathon path on AVP.

4 · Wire audio. Drop the .mp3 ambient loop on a looping audio source; attach object SFX to the matching props.

5 · Build to device.

  • Meta Quest 3 — Unity (URP) or Unreal via OpenXR, Android build, hand tracking + controllers. The smoothest engine target.
  • Apple Vision Pro — Unity PolySpatial (shared space) or a fully-immersive visionOS app; favor the mesh environment for performance.

Hackathon takeaways

  • Meshes are easy; splats are the risk. The .glb/.obj outputs drop straight into any engine. The .spz environment needs conversion and an experimental renderer — on Quest 3 budget it carefully, and on Vision Pro prefer a mesh environment unless you’ve already proven a splat renderer.
  • Lower the face count up front — regenerating is a one-line flag; re-decimating later is not.
  • Best polish/perf comes from this engine route; if you just need something running today across Quest 3 + Vision Pro, the WebXR route is faster to first pixel.

Questions? Reach us via the Contact page.

// BACK TO NEWS