Infinigen

Princeton's open-source procedural 3D world generator, where every asset is generated from scratch by randomized mathematical rules.

Infinigen was developed by Princeton's Vision and Learning Lab (Jia Deng's group), built on Blender, and open-sourced under the BSD license. The first version (CVPR 2023) generates natural outdoor scenes — terrain, plants, animals, and phenomena like fire, clouds, rain, and snow — with every asset's shape and texture generated from scratch by randomized mathematical rules, using no external assets at all, which allows infinite variation and automatic output of ground-truth depth, optical flow, segmentation, and normals. Infinigen Indoors (CVPR 2024) extended this indoors, using a constraint description language and a solver to place furniture and appliances, and can export to real-time simulators such as Omniverse for training embodied agents. Infinigen-Articulated (also called Infinigen-Sim, 2025) further procedurally generates articulated, simulation-ready objects, exportable as URDF, MJCF, or USD for use in MuJoCo and Isaac.

ExampleRunning the same indoor-generation rules with a different random seed produces a new room with a different layout, furniture shapes, and materials, complete with pixel-accurate ground-truth depth and segmentation, ready to use as visual training data.

Also called

Infinite Photorealistic Worlds using Procedural Generation, Infinigen Indoors, Infinigen-Articulated, Infinigen-Sim

Related

Procedural Generation · Synthetic Data · Simulation Assets · Articulated Object · Blender · Generative Simulation

Sources

Infinigen official site
Infinite Photorealistic Worlds using Procedural Generation (arXiv 2306.09310)
princeton-vl/infinigen (GitHub)

As of

2025-05