Reproduce “EntityBench: Towards Entity-Consistent Long-Range Multi-Shot Video Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ruozhen He, Meng Wei, Ziyan Yang, Vicente Ordonez
EntityBench offers a comprehensive benchmark for evaluating entity consistency in long-range multi-shot video generation. EntityMem, a memory-augmented system, improves character fidelity and cross-shot consistency, outperforming baselines in LLM-judged metrics.
Key insights
- Explicit, quality-gated entity memory management significantly improves character fidelity and cross-shot consistency, demonstrating more robust performance over long sequences and recurrence gaps compared to implicit approaches.
- There is a structural disagreement between DINOv2 embedding similarity metrics and LLM-based identity judgments; high embedding similarity does not always correlate with human-perceived entity identity consistency.
- Achieving high visual quality in individual frames is distinct from maintaining consistent entity identities across long, multi-shot narratives, underscoring the necessity of specialized, fine-grained evaluation like EntityBench.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.