WorldMemBench: A Comprehensive Benchmark for Memory of Video World Models
WorldMemBench tests video world-model memory across 260 static and dynamic cases.
WorldMemBench evaluates memory in video world models across 260 test cases, separating static retention of people, objects, and environments from dynamic event progression and state persistence. Controlled camera departures and multi-revisit trajectories compare static content with initial references and dynamic events with expected state transitions, alongside video-quality and camera-following diagnostics. Evaluation of nine models finds that strong static retention does not imply reliable dynamic evolution, and repeated revisits expose degradation after the first return. Human studies support alignment between the automatic metrics and human judgments.
- WorldMemBench has 260 cases covering static and dynamic world memory.
- Static tests cover identity, objects, and environment appearance and geometry.
- Dynamic tests cover event progression and state persistence across revisits.
- Nine models show strong static memory does not ensure reliable state evolution.
From the original article
1 min read at huggingface.co
As video models enable longer and more realistic interactive experiences, maintaining a world consistent with its history becomes essential. This requires both preserving previously observed content and ensuring coherent world evolution over time. World memory is a critical capability to evaluate alongside visual fidelity and video duration.
By Nan Zhou, Jianxiong Gao, Qiuyu Wang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.