image-blaster: Turn One Photo Into an Explorable 3D World in Under 5 Minutes
Every so often a repo lands that makes a whole discipline go quiet for a day. image-blaster is one of those. The pitch is almost rude in its simplicity: input an image, output a world, five minutes. Meshes with physics, a background splat, ambient audio — the works.
What it is
image-blaster is an open-source (MIT) “image-to-world” skillset for Claude. It isn’t a single model — it’s an orchestration layer. You give Claude your API keys, drop a picture into the input/ folder, and ask it to “blast” the image. Claude then drives a pipeline of specialized generative models and assembles the results into a coherent 3D scene.
From one photo, by default you get:
- A Gaussian splat (
.spz) of the static environment - 3D models (
.glb,.obj) of all the dynamic objects in the scene - Ambient looping sound plus object-specific physics SFX (
.mp3)
That’s a textured, explorable environment with audio — the kind of thing that’s historically taken a 3D artist days.
The model stack
The clever part is which tools it chains together:
- marble-1.1 (World Labs) — builds the explorable environment / splat
- hunyuan-3d (via FAL) — generates the individual 3D object meshes
- nano-banana — default image editor for source cleanup and clean plates
- gpt-image-2 — alternate image-edit provider
- elevenlabs-sfx — ambient and object-specific sound
Claude is the conductor; World Labs, FAL, and ElevenLabs are the orchestra.
It’s a Claude skill, not an app
The workflow is pure agent-native:
git clone https://github.com/neilsonnn/image-blaster
cd image-blaster
claude # install: curl -fsSL https://claude.ai/install.sh | bash
# give Claude your World Labs + FAL keys, drop an image in input/, say "blast it"Claude confirms each step with you as it goes. Because it’s a skillset rather than a locked binary, you get real control — Hunyuan parameters are exposed for face count (defaults to a sane 50,000 vs. the API’s 500,000), PBR materials, and generation type (Normal, LowPoly, or geometry-only). There’s even a bundled React viewer you can let Claude modify by un-ignoring /app.
Why it matters
Two things stand out. First, the output is production-usable — these aren’t toy renders, they’re assets you embed directly in Unity, Unreal, Godot, Blender, Maya, or a web app. Second, it’s a clean demonstration of where agentic tooling is heading: the value isn’t any single model, it’s Claude coordinating a half-dozen frontier APIs into one coherent artifact, confirming decisions with you along the way.
Video-game level concepts, your childhood bedroom, a film location scout’s reference, an architectural rendering — image-blaster turns any of them into a walkable scene before your coffee gets cold. The people who spent ten years learning Blender are right to stare at it in silence for a minute. Then they should go use it as a jumpstart.