What it does

image-blaster is a Claude skillset — essentially a prompt-and-tooling scaffold — that takes one image and farms out the work to specialized generative services. World Labs builds the explorable environment, FAL’s Hunyuan-3D handles object meshes, ElevenLabs generates ambient and physics-matched SFX, and various image models clean up source plates. The outputs are standard formats: .glb/.obj for objects, .spz gaussian splats for static scenery, .mp3 for audio.

The interesting bit

The orchestration is the product. Rather than building its own models, this is glue code with taste — it knows which API to call for which asset type, how to parameterize Hunyuan (face counts from 40K to 1.5M, PBR toggle, low-poly modes), and how to keep a human in the loop with Claude’s step-by-step confirmation. The “< 5 minutes” claim is in the README; your mileage will vary with API latency and queue depth.

Key highlights

• Outputs plug into Unity, Unreal, Godot, Blender, Maya, or Three.js without conversion drama
• Separates dynamic objects (meshes) from static environment (gaussian splat) — a sensible pipeline choice
• Includes a React viewer app that Claude can modify if you remove /app from .claudeignore
• Hunyuan parameter exposure is thorough: face count, PBR, polygon type, generation mode
• Requires API keys for World Labs and FAL; ElevenLabs and others likely need their own credentials too

Caveats

• README is thin on error handling, retry logic, or cost estimates across five paid APIs
• The “confirm each step with me” workflow suggests this still needs babysitting; not a fire-and-forget tool
• No topics, no tests visible, no CI — 4.4K stars but the repo surface is minimal

Verdict

Worth a look if you’re prototyping 3D environments and already pay for World Labs, FAL, and friends. Skip it if you need a standalone offline pipeline or predictable per-image pricing.