EntityBench: Towards Entity-Consistent Long-Range Multi-Shot Video Generation

Researchers introduced EntityBench, a benchmark of 140 narrative episodes and 2,491 shots for testing long-range multi-shot video generation.

The suite evaluates quality, prompt alignment, and cross-shot entity consistency across explicit difficulty tiers.

They also propose EntityMem, a memory-based method that stores verified visual references to preserve characters, objects, and locations.

Results show entity consistency degrades as recurrence gaps grow, while EntityMem performs best on character fidelity and presence.