Three-dimensional panorama of a dining room with a long table, chairs and windows in a castle

· Translated from the Spanish original

Artificial intelligence

Mundo turns a few photos of a room into a world you can explore and simulate

Mundo builds a 3D mockup of a place from photos, moves it with a simulator and shows it as video. It was trained on about 40,000 video pairs.

What happened

  • On October 23, 2026, a team with Yihang Luo as first author and Chen Change Loy among the signatories published Mundo on the Hugging Face blog, a world model that starts from photos or a video of a place.
  • Mundo rebuilds the scene as a 3D mockup made of boxes that approximate objects. A vision-language agent then corrects that mockup by comparing its renders with the original images.
  • A scene program defines physical properties and a simulator computes how objects move. A video model draws the result using the photos as context.
  • With that you can walk through the room from other angles, open a drawer, drop an apple or test a robot task.

Why it matters

  • Separating physics, which a simulator computes, from appearance, which a video model paints, can help objects obey physical laws, although the text does not quantify it.
  • For robotics and driving, recreating a real site from a few images allows testing tasks without building the physical scene.
  • The text gives no overall comparison against other systems with figures. It shows examples and a visual reconstruction comparison in a hotel room.

The number

40,000 pairs. Real or synthetic videos paired with their mockup version, used to fine-tune the video model.

Context

The authors explain that 3D detectors often miss objects or label them wrongly, which is why they brought in the vision agent as a reviewer. For training they gathered clips of real and synthetic interiors, driving, hand manipulation, single objects and video games, plus their own material, and fine-tuned an existing video model on 124-frame clips. The article includes a limitations section that anyone wanting to replicate the method should read.

What’s next

  • The blog points to a project page with more demos.
  • The reviewed text gives no date for code or weights.

Bottom line

Simulating real scenes is already visible in industry: Rivian bought Third Dimension to recreate real driving scenes in 3D. In video, SolarRenderer turns a video game’s structure into realistic video in 4 steps. To judge whether an agent explores well, a test bench invents worlds with false rules, and in hardware a micro-drone under one gram flies untethered for more than four minutes.

Sources

Edited by Rodrigo Cornejo. How we select and verify: who writes these notes.

Related notes

← All notes