image-blaster: one photo becomes a walkable 3D world

Turning one photo into a 3D scene you can walk through used to mean wiring up several generation services yourself: one for the environment, one for object models, one for sound, each with its own account and API. image-blaster is an open-source Claude skill pack by indie developer neilsonnn (MIT, TypeScript) that turns that pipeline into a manual your coding agent follows: say “blast it” and it runs photo analysis, clean plate, environment generation, object modeling and sound effects in order. The WeChat account 极客之家 featured it on October 7 with a full hands-on run; as of October 8, 2026, the repo created in April 2026 has 10,334 stars.

Core features

• Three asset types from one photo: the static parts of the picture (walls, floor, sky) become a .spz Gaussian-splat scene you can roam; movable things — chairs, tables, crates — get their own .glb and .obj models; an ambient loop plus per-object collision sounds come as .mp3.
• An agent runs the whole pipeline: the repo is really a set of SKILL.md files and scripts under .claude/, built for Claude Code, and 极客之家 confirmed it works inside Codex too. Ask it to confirm each step before running, and strike anything off the modeling list you do not want.
• Four generation models underneath: environments come from World Labs Marble (marble-1.1), object meshes from Tencent Hunyuan 3D, sound from ElevenLabs — the latter two billed through the FAL platform. Image edits that erase objects default to nano-banana and can switch to gpt-image-2.
• Clean plates prevent ghosts: the agent erases the listed objects from the photo first, then generates the environment from the empty plate. Skip that step and the erased objects linger as ghosts in the 3D scene — a concept borrowed from film post-production.
• Numbered outputs you can retrace: the source image is file 0 and is never overwritten; every generated image and request log is kept. If one step goes wrong, rerun just that step instead of paying for the whole run again.
• Modeling knobs: face count defaults to 50,000 (Hunyuan accepts 40,000–1,500,000), with LowPoly and Geometry modes, triangle or quadrilateral faces, and PBR materials on by default.

Typical use cases

• Game level concepting: run one mood-board photo and walk the resulting scene minutes later — the README claims under 5 minutes from photo to meshed environment.
• Film scouting and architecture previews: turn reference photos into a roamable space to check sightlines and movement before committing.
• Robot environment drafts: sketch a navigable scene quickly for concept work — read the caveats below before using it for simulation.

Quick start

You need API keys for both World Labs and FAL, both metered per call. The commands below come from the official README:

git clone https://github.com/neilsonnn/image-blaster
cd image-blaster
claude

Drop a photo into input/, give Claude the two keys, then say “blast it and confirm each step with me”. It analyzes the photo and lists what it plans to model, waits for your confirmation, then runs the rest on its own.

Summary

image-blaster suits game developers, indie hackers and 3D hobbyists who want to see a concept scene quickly, especially those already living in Claude Code. It is MIT-licensed and written in TypeScript. Know these before spending money:

• The back side is invented: anything the camera did not see is the model’s guess — show it one wall of a room and there is no basis for what is behind it, so the result approximates your room rather than reproducing it.
• Physics is uncalibrated: colliders are rough and object size and weight are not calibrated; 极客之家 states outright that the output is not ready for robot simulation training.
• Every step is a paid API: Marble, Hunyuan 3D, ElevenLabs and nano-banana are all closed models, so one run draws on both your World Labs and FAL balances — failed experiments cost too.
• The repo has been dormant for five months: the last commit landed on May 15, 2026; ten thousand stars in half a year, but no maintenance, so changes in the underlying model APIs are yours to patch.
• If you only want an agent inside an existing Blender modeling workflow, the previously covered mcp-for-blender takes the other route: it drives the tools you already use instead of generating a world from a photo.