Orbia: A New Benchmark for Evaluating 3D Consistency in AI-Generated Interactive Video Worlds
arXiv:2610.20693 (cs.CV) โ Submitted 14 Oct 2026
Authors: Dongyue Lu, Guian Fang, Hur Lim, Yonggang Wu, Yanrong Wang, Hongshan Chen, Shengju Qian, Yue Liao, Wei Chow, Lingdong Kong, Tianxin Huang, Yu Yang, Hongji Yang, Junchao Huang, Ao Liang, Yihao Wang, Zihan Wang, Rong Li, Hanlin Chen, Xin Wang, Wei Tsang Ooi, Mike Zheng Shou, Shuicheng Yan
Abstract
Video generators increasingly serve as interactive worlds for exploration and simulation, yet realistic and visually appealing frames do not necessarily depict a consistent 3D world. Existing benchmarks assess selected geometric properties or use vision-language models to judge memory and consistency, leaving systematic 3D evaluation underexplored.
The authors introduce Orbia, the first benchmark dedicated to a systematic and comprehensive evaluation of 3D consistency in interactive world models. Through controlled exploration and revisit, the evaluation is organized into three dimensions:
- Input preservation โ Registered RGB-D references and annotated anchors support appearance, geometry, and spatial stability measurements.
- Generated-content persistence โ Revisits to regions absent from the input image are compared with their first generated observations to test whether models preserve what they create.
- 3D self-consistency โ Cross-view depth and feature correspondence, regional scale agreement, and held-out Gaussian rendering test whether generated views describe the same 3D scene.
Why 3D Consistency Matters in 2026
As world models and video generators mature in 2026, they are being deployed as explorable environments for robotics simulation, embodied AI, and interactive content creation. A generator that produces photorealistic frames can still fail catastrophically as a "world": walls shift between views, objects morph when revisited, and depth relationships collapse โ silently breaking downstream planning and control. Orbia addresses this gap by treating 3D consistency as a first-class, measurable property rather than a byproduct of visual fidelity.
Evaluation Dimensions in Detail
Input Preservation
Using registered RGB-D references and annotated anchors, Orbia measures whether generated frames faithfully retain the appearance, geometry, and spatial layout of the input scene. This dimension isolates how well a model respects its conditioning signal.
Generated-Content Persistence
When a model explores regions outside the input image, it invents new content. Orbia revisits those regions and compares them with the first generated observation, testing long-horizon memory as well as the fidelity of self-generated content.
3D Self-Consistency
This dimension uses cross-view depth and feature correspondence, regional scale agreement, and held-out Gaussian rendering to verify that generated views describe one coherent 3D scene. Gaussian-splatting-based rendering provides a differentiable, geometrically grounded check on scene plausibility.
Key Takeaway
Orbia provides minute-long, controlled exploration protocols that stress-test world models across appearance, memory, and geometry. It offers the community a rigorous, reproducible standard for the emerging class of interactive video generators โ a necessary foundation as these systems move from demo reels into simulation-critical applications.
via ArXiv CV