SJTU's StrideWorld Benchmarks 15 Interactive Video World Models, Finds All But One Deform Scenes by Up to 7.3x
Summary
Shanghai Jiao Tong's StrideWorld benchmark fixes scene references before evaluation and catches drift that older tests miss: across 54,600 videos from 15 open interactive video world models, all but one deformed scenes with offsets up to 7.3x those of rigid rendering. Only two depth-scaled command models stayed close to stride-predicted positions on composed routes; most stalled or landed off target.
Originally reported by huggingface.co
Read the original article →Original headline: SJTU's StrideWorld Benchmarks 15 Interactive Video World Models, Finds All But One Deform Scenes by Up to 7.3x