Research
WanderBench: A New Benchmark for Video-Conditioned World Models Exploring Dynamic Worlds
WanderBench evaluates video-conditioned world models on exploring dynamic scenes from real and AI-generated videos, using 745 source videos across three difficulty tiers and 24 human-validated sub-metrics. Benchmarking eight models shows that longer clips preserve static scenes and visual quality but degrade camera control, source dynamics, and revisit consistency. The authors also propose WanderLand, a training-free conditioning framework that improves five models on every evaluated dimension, including up to 8.3 points in source-dynamics fidelity.
world modelsvideo-conditioned generationbenchmarking