Orbia: Do Generated Videos Form a Consistent 3D World?

Orbia. Do Generated Videos Form a Consistent 3D World?. By Dongyue Lu and twenty-two others, from the National University of Singapore, LightSpeed Studios, and six other institutions. Published on the fourteenth of October. Its code repository is worked out: about one hundred source files, with a readme and a licence.

Video generators are now used as interactive worlds, for exploration and for simulation. But frames that look realistic and appealing do not necessarily depict a consistent 3D world. Existing benchmarks either assess selected geometric properties, or ask vision-language models to judge memory and consistency. Systematic 3D evaluation has been left underexplored.

Orbia is presented as the first benchmark dedicated to a systematic and comprehensive evaluation of 3D consistency in interactive world models. It works through controlled exploration and revisit: a model is driven away from a place and brought back. The evaluation is organised into three dimensions. Input preservation asks whether the model keeps what it was given. Generated-content persistence asks whether it keeps what it made up itself. And 3D self-consistency asks whether its views all describe one scene.

Each dimension has its own instruments. For input preservation, registered RGB-D references and annotated anchors support measurements of appearance, geometry and spatial stability. For generated-content persistence, the benchmark looks at regions absent from the input image, and compares a revisit of such a region with the first time the model generated it. For 3D self-consistency, cross-view depth and feature correspondence, agreement in regional scale, and held-out Gaussian rendering all test whether the generated views describe the same 3D scene. Minute-long exploration and revisits push the test toward long-term consistency.

The construction and annotation pipeline draws on public datasets, scenes the authors captured themselves, and simulation. The result is seven hundred and twenty instances, spanning indoor and outdoor environments and six exploration-and-revisit protocols. Evaluation covers twenty-seven models, conditioned variously on text, on actions and on camera, each connected through an adapter for its native interface.

What the evaluation is built to surface is the opening concern: frames that look realistic and visually appealing need not depict a consistent 3D world. Orbia separates the ways that can go wrong. A model can fail to preserve the input it was given. It can fail to keep what it generated itself, so that a region looks different on revisit than when it first appeared. Or its views can fail to describe the same 3D scene. Each of these is measured on its own, across indoor and outdoor scenes, and over minute-long exploration as well as short revisits.

Orbia is offered as a systematic basis for two jobs: diagnosing 3D inconsistency in generated video, and developing video world models whose worlds persist. The title's question, whether generated videos form a consistent 3D world, is turned into something that can be measured from several directions at once.