# Larracat

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Ondrej Jakub Chromy](https://cerebralvalley.ai/u/_OJ_)
- **GitHub:** https://github.com/OJrogue/movable-reality
- **Demo video:** https://youtu.be/S1o0vRiPR7w
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4582

Movable Reality turns one video of a room into a 3D room that you can adjust.

A 3D Gaussian Splatting scan looks convincing, but it is one frozen block of millions of splats. You can look around it, but you can't pick up the chair. Movable Reality makes each object in the room its own splat, so you can move it, turn it, remove it and bring it back.

One command runs the whole pipeline. It takes a video of a slow walk through the room, plus the focal length and ceiling height, and needs no depth data, floor plan or hand labels. Muse Wonder watches the video and lists the objects. SAM 3.1 tracks each one through the video. MapAnything and COLMAP recover the cameras, gsplat trains the scene, and the masks are lifted onto individual Gaussians to separate the objects.

The result opens in a browser. You can drag objects with the mouse or give the agent a command such as "Delete the ukulele and put a ceramic vase on the dresser." The agent plans the edit with real dimensions and carries it out, and a second model checks the result in a screenshot. Every edit can be undone. You can also add furniture from a library, measure distances, see the spots the video barely covered, and compare scanned objects with SAM 3D Objects versions.

The demo has two rooms rendered in Blender:
- Bedroom: 54 s of video, 10 movable objects, 1.24 M splats, 36.9 dB on held-out frames of the same walk.
- Living room: 122 s of video plus a 10 s top-up, 39 movable objects, 3.37 M splats, 28.4 dB on a separate test walk.

It is a working demo on two synthetic rooms and a good place to start for research or a project. The code is open under MIT and includes the separated object splats, a single JSON scene file, agent tools, and Blender rooms with ground truth to measure changes against. It is built for people working on object-level splats, room editing, or agents that act on 3D scenes.

## More from Meta Global AI Developer Hackathon

- [Thoth](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4579)
- [MEEMAW](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4580)
- [LightSeat](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4581)
- [Kinetix Labs](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4583)
- [FullCircle ComXpress](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4584)
- [Ferrum](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4585)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4582. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt