From video generation to world modeling: a trajectory and memory-based evaluation framework
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Recent advances in generative video models have led to the emergence of video world models, which aim to generate coherent scenes with controllable camera motion and persistent scene structure. However, systematic evaluation of these capabilities remains limited. This project proposes an evaluation framework for analyzing video world models along two key dimensions: camera trajectory control and appearance persistence. For camera motion evaluation, camera trajectories are recovered from generated videos using DROID-SLAM and compared with idealized trajectories using quantitative metrics such as RMSE and angular error. To assess scene memory, an appearance persistence benchmark is introduced, where a camera moves away from an object and later returns. A VQA-based evaluation protocol using the Gemini multimodal model is employed to automatically score object consistency. Compared with existing benchmarks that primarily focus on perceptual quality or short-term temporal consistency, this work provides a more structured and physically grounded evaluation of camera motion and scene persistence. The framework is model-agnostic and enables systematic diagnosis of motion and memory failures, facilitating the development of more reliable video world models.
Fields of Study & Subjects
Generative Adversarial Networks and Image Synthesis
Multimodal Machine Learning Applications
Advanced Vision and Imaging
Publication details
- Journal / source
- DR-NTU (Nanyang Technological University)
- Publication date
- 1 January 2026
- Language
- EN