Introduction

One photo in. Meshes, a Gaussian splat room, and physics SFX out — under five minutes if the APIs cooperate. That is the bet behind image-blaster, an open-source Claude skillset from Neilson Koerner-Safrata (neilsonnn) on the World Labs design team.

Primary artifact: github.com/neilsonnn/image-blaster (MIT). Created 21 Apr 2026 MYT; latest commit on main is the 15 May 2026 license drop. This is not a same-week weight dump — it is a still-runnable orchestration pipeline that World Labs also showcases. The repo is open source; the generators behind it are API products (World Labs Marble, FAL Hunyuan 3D, ElevenLabs SFX, plus image-edit providers). No local open weights ship here.

The Magic: Photo → Engine-Ready World Kit

Drop an image into input/, tell Claude to blast it, and the skill chain aims to emit:

  1. Dynamic object meshes — .glb / .obj via Hunyuan 3D on FAL
  2. Static environment splat — .spz via World Labs marble-1.1
  3. Audio — ambient loop + object physics SFX as .mp3 via ElevenLabs

The workflow is agentic, not a single monolithic model: Claude skills clean the plate, uncover / separate dynamic props, call Marble for the explorable splat world, mesh the interactables, and attach sound. World Labs’ own showcase frames it as Marble + Claude skills + fal for 3DGS environments, meshes, interactive physics objects, and SFX from one image.

Stack (what actually runs)

PieceRole

marble-1.1 (World Labs)

Explorable Gaussian-splat environment

nano-banana

Default image edit (cleanup, clean plates, object refs)

gpt-image-2

Alternate image edit when the edit skill prefers it

hunyuan-3d (via FAL)

Textured / low-poly / geometry meshes

elevenlabs-sfx

Ambient + object-specific SFX

Hunyuan knobs the README documents (and image-blaster defaults differently from the raw API):

  • --face-count — allowed 40000–1500000; image-blaster default 50000 (Hunyuan API default is 500000)
  • --enable-pbr — default true
  • --generate-type — Normal (textured) | LowPoly | Geometry (white geo); default Normal
  • --polygon-type — triangle | quadrilateral for LowPoly; default triangle

Keys live in .env-style config: WORLD_LABS_API_KEY + FAL_KEY. Claude Code / Claude CLI drives the session; a Bun workspace under app/ hosts a React viewer (gated by .claudeignore until you unlock it for skill edits).

Pro Tip

Budget the mesh tier, not the splat. Keep Marble for the static shell you walk, and push only true interactables through Hunyuan at 50000 faces with PBR on. Hero props can bump --face-count; background clutter stays LowPoly or Geometry. That split is what makes Unity / Unreal / Godot import sane instead of dumping a half-million-face soup into every prop.

Hands-on: Blast It Today

git clone https://github.com/neilsonnn/image-blaster
cd image-blaster
# install Claude CLI if needed:
# curl -fsSL https://claude.ai/install.sh | bash
claude

In the Claude session: hand over your World Labs and FAL API keys, drop a photo into input/, then paste:

blast it and confirm each step with me

Confirm each step so you can reject a bad plate or mesh before the next API burn. When assets land, embed them under engine / DCC / web paths the README calls out: Unity, Unreal, Godot; Blender, 3ds Max, Maya; Three.js / Electron. World Labs’ showcase also stresses export into Unity, Unreal, and Three.js.

Why It Matters for Game Pipelines

Most image→3D tools stop at a single hero mesh. image-blaster targets a level-kit shape: splat room for presence, discrete meshes for interaction, SFX for feedback — the three buckets a greybox pass actually needs. Because orchestration is open MIT and the heavy lifting is billed APIs, you can fork the skill graph, swap an edit provider, or retarget export without waiting on a closed world-model SDK.

Age check, stated plainly: repo born mid-April 2026, last main push mid-May 2026. Still useful as a reference pipeline and a runnable Claude skillset if your keys and Claude CLI are current — just do not pitch it as a brand-new Oct 2026 drop.

Conclusion

image-blaster is a Claude-orchestrated image-to-world kit: one photo → .glb/.obj dynamics, .spz static env, .mp3 SFX, aimed at sub-five-minute turnarounds and engine/DCC/web embed. Maker Neilson Koerner-Safrata; MIT skillset on GitHub; generation via Marble 1.1, FAL Hunyuan 3D (50k-face default), and ElevenLabs — not open local weights. Clone it, feed keys, blast a level concept photo, and see whether the mesh/splat/SFX split survives your first Unity or Godot import.

Source: neilsonnn/image-blaster · World Labs showcase

—Titus