Video world models struggle with long camera exploration over time
WorldProbe: When Video World Models Fall Apart under Long-Horizon Spatial Exploration
Computer Vision and Pattern Recognition
Summary
Video world models try to predict what a scene will look like in the future as a camera moves around. The authors found that these models work well for short movements but have trouble when the camera explores a large area for a longer time. They created WorldProbe, a test system that checks if the models can handle long and complex camera actions. Their tests show that while recent models keep images clear, they often lose track of the bigger, long-term scene details and consistency.
What this means in practice
- •For autonomous vehicle developers: Test video-based environment prediction systems over long drives with dynamic camera paths to detect long-term inconsistencies.
- •For augmented reality engineers: Improve testing of AR systems that rely on persistent world models by measuring how well visual worlds hold up during extensive user movement.
Authors
Minqi Wang, Jie Peng, Hejun Dong, Ruchang Yao, Puyi Wang, Siyuan Huang, Jihai Zhang, Yafu Li, Qi Dou, Xiaoye Qu, Yu Cheng
Abstract
Video world models generate future visual observations from camera or action inputs, with the goal of synthesizing persistent 3D worlds that can be explored across time and space. Existing benchmarks typically evaluate short interactions or rely on camera trajectories fixed before generation. The performance of video world models under long-horizon exploration with extensive camera motion remains under-explored. We introduce WorldProbe, a diagnostic benchmark that uses a goal-conditioned multimodal agent to adapt camera actions online during generation. WorldProbe comprises 164 test cases across four probing families. Across 10 state-of-the-art models, we generate more than 1,600 rollouts totaling approximately 800 minutes, with horizons up to 180 seconds. Our evaluation across four dimensions reveals both the degradation of local measures and the emergence of long-range inconsistencies. We find that recent models sustain visual and content quality for longer, while long-range memory and evolution remain weak. These results expose a gap between locally stable generation and persistent world consistency that is largely missed by existing benchmarks. Code and benchmark resources are available at https://github.com/Untitled-W/WorldProbe.