Engineering & AI
Image Blaster: draft a 3D scene from one image
Image Blaster turns one reference image into a walkable 3D scene so you can test the layout and sense of space before building a detailed version.
Summary
- Use a single reference image to test an idea in a walkable 3D draft with object models and sound.
- Every generation step runs on paid services (World Labs, FAL); a full run with five objects costs about USD 5, before sound.
- Whatever the photo does not show is invented by the AI, so the result suits concepts and drafts, not a replacement for a real 3D scan.

You have a reference image and want to walk through the scene before spending time modeling every object. Image Blaster uses Claude Code, World Labs, and FAL to turn that image into a walkable 3D draft with object models, ambient audio, and sound effects, so you can decide which scenes are worth developing.
Based on the source code at commit 4acb43b (15 May 2026), reviewed on 19 November 2026; we read the code but did not run it. HubCom has not used this tool in a client project.
What does it solve?
- Building a 3D scene is a 3D artist's job. The environment, the objects and the sound each come from a different tool.
- AI models exist for each part, but separately. World Labs builds environments, Hunyuan3D builds objects, ElevenLabs makes sound; each has its own API and format.
- The repo chains them into one workflow run by Claude Code: analyse the image, separate the objects, generate each part and save it locally. The README says an image can become a meshed 3D environment in under five minutes. HubCom has yet to time the full workflow. Image Blaster README
What it does
- It trains no model of its own. The project is a set of skills, subagents and scripts that make Claude Code call existing services in the right order.
- Image analysis: Claude describes the scene, lists each separable object and waits for you to approve the list.
- Clean plate: removes the approved objects from the image with Nano Banana 2 (default) or GPT Image 2.
- 3D environment: World Labs Marble 1.1 builds the space from the clean plate, as a Gaussian splat (
.spz) plus a collision mesh. - Per-object 3D models: isolates each object on a white background, then models it with Hunyuan3D v3 (default) or Meshy v6.
- Sound: ElevenLabs creates a looping ambience and impact sounds for each object.
- Local viewer: a React and Three.js app that runs on your machine to view and arrange the results.
- "Wildcard" skill: calls any model on FAL, but asks you to confirm before any paid request.
The project is MIT-licensed, with more than 12,500 stars on GitHub.
When to use it
- Use it to block out a 3D draft fast: game level concepts, film location ideas, a virtual showroom.
- Use it when a team needs many simple 3D models from existing photos, to refine later in Blender or a game engine.
- Use it to learn how to organise skills, subagents and hooks in Claude Code for a multi-step workflow.
- Don't use it to present a real home, hotel room or shop to customers. Anything outside the photo is drawn by the AI and may not match reality.
- Don't use it when you need dimensionally accurate models for manufacturing or 3D printing. A model built from one photo has no real measurements.
Caveats
- Every step is a paid request. Prices below are as listed on 19 November 2026.
- 3D environment (World Labs Marble 1.1): 1,580 credits, about USD 1.26 per world from a regular photo.
- Image edit (Nano Banana 2): USD 0.08 per image at 1K resolution.
- 3D model (Hunyuan3D v3): USD 0.375 per model, plus USD 0.15 for PBR materials and USD 0.15 for a custom face count; the skill enables both by default.
- Cost example: a scene with five objects costs about USD 5, before sound. Every regeneration costs again.
- World Labs can bill later. A request that starts on a low balance still runs; the overage is invoiced the next month, even with auto-refill off.
- One-shot mode runs many subagents in parallel: one 3D model per object and four impact sounds per object. Cost grows with the number of objects.
- API keys: the README suggests pasting keys into the chat so Claude writes
.env. Writing.envyourself is safer, because the keys never pass through the conversation. - The repo ships permissions and hooks.
.claude/settings.jsonpre-approves commands such asnode .claude/scripts/**andbun install, and runs a script when a session starts. Read it before you trust the folder. - Images made with Nano Banana 2 carry a SynthID watermark. See our ai-meta-cleaner review on watermarks and AI labels.
- Little recent activity: 72 commits between 21 April and 15 May 2026, no release yet. Pin the commit you use.
Trying it out
Before handing it to an AI agent, prepare:
- A source image: sharp, wide, with objects clearly apart. Your own photo, or one you have rights to use.
- Two API keys, World Labs and FAL, each account topped up with a small amount to cap the cost of the trial.
- The list of objects that need 3D models, to approve quickly at the analysis step.
- Where the results go: Three.js, Unity, Unreal, Godot or Blender, so you know which formats you need.
- The machine: Claude Code, Node.js, Bun, and ffmpeg if you want sound post-processing.
Rollout order
- Read
.claude/settings.jsonand.claude/hooks/before opening the project in Claude Code. - Create
.envyourself from.env.example, withWORLD_LABS_API_KEYandFAL_KEY. - Put one image in
input/and ask Claude to go step by step, waiting for your confirmation after each step. - Approve the object list, dropping anything that does not need a 3D model.
- Check the clean plate before generating the world; regenerating the world is the most expensive step.
- Generate models and sounds for the approved objects, then review them in the local viewer.
Common pitfalls
- Running one-shot on the first try. Every paid step runs back to back, with no stop to fix a mistake.
- Leaving objects in the clean plate. They end up both in the environment and as separate models.
- Trusting the environment to match the real place. Behind, beside and around corners, the AI is guessing.
- Running on Windows without Git Bash or WSL. The hooks are bash scripts and the port check uses
lsof. - Missing Bun or ffmpeg. The README does not mention them, but the viewer needs Bun and sound processing needs ffmpeg.
- Committing
worlds/to Git..spz,.glband audio files are heavy; keep them out of the repo or use Git LFS.
Checking the result
- Every approved object has a model file in
worlds/<name>/output/<object>/that opens in Blender or a GLB viewer. - The world has a
.spzfile and a collision mesh inoutput/world/, downloaded locally, not provider links. - The local viewer shows everything: environment, objects and sound.
- Actual cost matches the estimate: check the World Labs and FAL usage pages after the first run.
Out of scope
Comparing 3D model quality, bringing Gaussian splats into each game engine, and content ownership under each provider's terms. See the project source.
Terms in this article (8)
- image-blaster
- An open-source Claude Code skillset that turns one image into a 3D environment, object models and sound effects.
- Gaussian splat
- A way to represent a 3D scene with millions of soft coloured points, which looks close to a photo as you move the camera.
- World Labs Marble
- World Labs' model that builds walkable 3D environments from images, text or video.
- Hunyuan3D
- Tencent's model that creates a textured 3D model from a single image.
- FAL
- A service that runs many AI models for images, video, audio and 3D through an API, billed per request.
- Clean plate
- The scene image with the objects to be modelled removed, used to build the empty environment.
- GLB
- A common 3D model file format that holds geometry, textures and materials in one file.
- Subagent
- A secondary AI agent that Claude Code gives a separate task, able to run in the background alongside others.
Frequently asked questions
Is image-blaster free?
The code is free under the MIT licence, but every generation step calls a paid service. A scene with five objects costs about USD 5 at prices listed on 19 November 2026, before sound.
Does an image-blaster world match the real place?
No. World Labs Marble builds the space from one image, so everything the image does not show is drawn by the AI. Use it for concepts and drafts, not to present a real location.
Which game engines can use image-blaster's output?
Object models are GLB files that open in Blender, Unity, Unreal, Godot and Three.js. The environment is a Gaussian splat (.spz), which needs a tool that supports the format; the bundled viewer uses Three.js.
Sources
- image-blaster, GitHub
- World Labs API pricing, World Labs
- Hunyuan3D V3 Image to 3D, fal
- Nano Banana 2 Edit, fal