For those who don’t know the tool: image-blaster sits between concept and rough layout, using Claude to orchestrate Marble from World Labs and write splat environment files you can move into DCCs and engines.
The headline feature is the splat
image-blaster takes a single reference image and, by default, generates a Gaussian splat of the static environment as an .spz file. This splat is a static environment output, separate from any object meshes or audio it may also generate. It runs as a Claude driven workflow where you put an image into the input folder and ask Claude to blast it while confirming each step.
That .spz environment output targets the fast part of production where you need space, scale, and a place to put cameras. If your team already tracks experiments for gaussian splats, this lands squarely in that bucket, with the advantage that the output arrives as a file on disk rather than a locked viewer.
There is also a time claim of under five minutes to go from image to a fully meshed 3D environment. Treat that as a marketing claim and validate it against your own network, machine, and API limits before you build expectations around it.
Marble, SPZ, and what you actually get
image-blaster lists marble-1.1 as the world generation model it uses for the explorable environment. In practice, that means the environment representation you get from the Marble side is an SPZ splat file. SPZ is Marble’s native splat format, optimised for file size, while PLY is an uncompressed format compatible with more Gaussian splat software packages.
If you plan to wire this into a pipeline, the file format matters more than the demo. SPZ tends to travel well when you stay inside the same ecosystem, and PLY tends to travel well when you want broader compatibility. World Labs also documents a common gotcha: default generated worlds use an OpenCV coordinate system, while many DCC applications use an OpenGL coordinate system, so correcting for that by scaling the Y and Z axes by minus one while keeping X unchanged is necessary. We learned that the hard way.
Splat assets exported by the API use arbitrary model units, not meters, and points to semantics metadata values that can be used for metric scale and ground alignment. That is the kind of detail you want early, because nothing burns time like a “looks fine in the viewer” asset that arrives in the wrong coordinate frame once you ingest it. And, let’S be honest, generated stuff lacks in consistency, and everybody knows. Plan for that!
The workflow is orchestration, not a single model
image-blaster frames itself as a set of skills that orchestrate multiple models. Marble handles the environment generation path that results in the SPZ splat for the static environment. The repository has other models it can involve for cleanup and object generation, but you can treat those as optional if your goal is strictly to generate splats from images.
You run the Claude CLI, provide API keys for World Labs and fal, then feed an image and guide the process via prompts. The repository even includes a development note about removing the app folder from the Claude ignore file if you want Claude to change the React viewer. That means the output can become part of a toolchain rather than a one off experiment. You can keep the same wrapper, swap inputs, and iterate, which is usually what artists actually want when they say “make it usable.”
Where splats meet real tools
image-blaster says you can embed it under the assets of a game engine, DCC software, or a web app. It explicitly names Blender as a target DCC and calls out Three.js for web apps.
image-blaster is open source under the MIT license, but the environment generation step relies on paid APIs. World Labs API billing runs on credits purchased through the World Labs Platform, and credits purchased for the Marble app cannot be used with the API. The API pricing page has a fixed rate of 1.00 USD per 1,250 credits, with a minimum purchase of 6,250 credits or 5.00 USD.
Pipeline advice that saves you from yourself
Splat workflows reward quick iteration and punish sloppy ingestion. Lock down where SPZ files live, how they are versioned, and who converts them, especially if you plan to move between coordinate systems or into other formats. Also, test new tools and innovations before use in production. Verify coordinate orientation, scale, file integrity, and viewer behavior before anyone depends on the output for layout, virtual scouting, or client reviews.
https://github.com/neilsonnn/image-blaster
https://www.worldlabs.ai/
https://docs.worldlabs.ai/marble/export/specs
