Image Blaster is an image-to-world skillset for Claude. It reads a source image, decides which objects are worth meshing, builds an explorable environment, and generates ambient and object-specific sound — usually in under five minutes.
What you need
- A clone of the Image Blaster repository
- Claude Code, installed with
curl -fsSL https://claude.ai/install.sh | bash - A World Labs API key from platform.worldlabs.ai
- A FAL API key from fal.ai
- Node.js and Bun for the bundled scripts and the viewer
Copy .env.example to .env and fill in WORLD_LABS_API_KEY and FAL_KEY.
Run a blast
- Open a terminal and
cd image-blaster. - Run
claudeand share both API keys when asked. - Put an image into the
input/directory. - Ask Claude to
blast it and confirm each step with me.
Claude stages a project under worlds/<slug>, analyzes the source image, confirms the object list with you, then generates the world, the meshes, the clean plate and the sound effects.
What comes out
By default, an Image Blaster blast produces:
- 3D models (
.glb,.obj) of every dynamic object - A Gaussian splat (
.spz) of the static environment - A clean plate with the foreground objects removed
- Ambient looping sound and object-specific physics SFX (
.mp3)
Everything is written to disk under worlds/<slug>. Provider URLs in the JSON sidecars are provenance and resume metadata only, so generated assets keep working offline.
Watch it in the viewer
Image Blaster ships a React and react-three-fiber viewer in app/. From the repository root run bun install && bun run dev, then open the printed URL to walk the world, inspect each mesh and play the audio.
Tuning the meshes
Image Blaster’s image-blast-3d skill passes Hunyuan parameters straight through:
| Parameter | Values | Default |
|---|---|---|
--face-count | 40000–1500000 | 50000 |
--enable-pbr | true / false | true |
--generate-type | Normal / LowPoly / Geometry | Normal |
--polygon-type | triangle / quadrilateral | triangle |
Choose LowPoly when a game engine needs a lighter mesh, or Geometry when you only want a white geometry pass.
Taking it further
- Swap the image-edit provider between
nano-bananaandgpt-image-2. - Use the
image-blast-wildcardskill to run almost any FAL model after confirming it. - Change
--provideron the 3D skill to try Meshy instead of Hunyuan.
When something needs fixing, identify the right skill and rerun it rather than editing generated files by hand — Image Blaster’s skills keep project state and provenance consistent.