Einsia Physics Benchmark Exposes Video AI Gaps

This is a valuable corrective to video benchmarks that reward realism without testing whether anything obeys nature. The score is sobering, though it should be read as a diagnostic of observable physics—not a universal measure of world-model intelligence.

• –Uses task-specific measurements instead of relying only on human or vision-language-model judgments
• –Separates visual continuity from physical correctness, a crucial distinction for robotics and planning
• –Seedance 2.5 leads overall, but 57.76/100 leaves substantial room for improvement
• –Covers physics beyond mechanics, including reflection, melting, fluids, and electromagnetism
• –Observability failures can suppress physical scores, so passing the benchmark still does not prove full physical consistency