EntityBench: Towards Entity-Consistent Long-Range Multi-Shot Video Generation
News Release Summary
Generating a long video story is not just a matter of making individual shots look good: characters, objects, and locations must remain recognizable when they return much later. EntityBench introduces a benchmark for measuring that problem across 140 narrative episodes and 2,491 shots, with explicit schedules for the entities that should recur. Its evaluation separates shot quality, prompt adherence, and cross-shot consistency, while avoiding consistency credit for incorrect appearances. The paper also presents EntityMem, a baseline that keeps verified visual references for each entity in persistent memory. The experiments show that existing methods lose consistency as recurrence gaps grow, whereas the memory-based approach achieves the strongest character fidelity among the compared methods.
abstract
Multi-shot video generation extends single-shot generation to coherent visual narratives, yet maintaining consistent characters, objects, and locations across shots remains a challenge over long sequences. Existing evaluations typically use independently generated prompt sets with limited entity coverage and simple consistency metrics, making standardized comparison difficult. We introduce EntityBench, a benchmark of 140 episodes (2,491 shots) derived from real narrative media, with explicit per-shot entity schedules tracking characters, objects, and locations simultaneously across easy / medium / hard tiers of up to 50 shots, 13 cross-shot characters, 8 cross-shot locations, 22 cross-shot objects, and recurrence gaps spanning up to 48 shots. It is paired with a three-pillar evaluation suite that disentangles intra-shot quality, prompt-following alignment, and cross-shot consistency, with a fidelity gate that admits only accurate entity appearances into cross-shot scoring. As a baseline, we propose EntityMem, a memory-augmented generation system that stores verified per-entity visual references in a persistent memory bank before generation begins. Experiments show that cross-shot entity consistency degrades sharply with recurrence distance in existing methods, and that explicit per-entity memory yields the highest character fidelity (Cohen's d = +2.33) and presence among methods evaluated.
citation
@inproceedings{he2026entitybench,
title = {EntityBench: Towards Entity-Consistent Long-Range Multi-Shot Video Generation},
author = {He, Ruozhen and Wei, Meng and Yang, Ziyan and Ordonez, Vicente},
year = {2026},
booktitle = {Conference on Neural Information Processing Systems. NeurIPS 2026},
url = {https://arxiv.org/abs/2605.15199},
}
automatically generated questions, main contributions and limitations of this paper
Questions this paper helps answer
- What does EntityBench measure? EntityBench evaluates whether multi-shot video generators keep characters, objects, and locations consistent across long narrative sequences.
- How large is the benchmark? It contains 140 episodes and 2,491 shots, with scenarios reaching 50 shots and recurrence gaps of up to 48 shots.
- How is consistency evaluated? The suite separately measures intra-shot quality, prompt-following alignment, and cross-shot consistency, using a fidelity gate so only accurate entity appearances count toward consistency.
- What is EntityMem? EntityMem is a memory-augmented baseline that stores verified visual references for each entity before generation and uses them throughout the sequence.
Main contributions
- EntityBench supplies per-shot schedules for characters, objects, and locations drawn from real narrative media.
- The benchmark covers easy, medium, and hard tiers with up to 13 recurring characters, 8 locations, and 22 objects in an episode.
- Its three-pillar evaluation separates image quality, prompt alignment, and long-range entity consistency rather than collapsing them into one score.
- EntityMem provides a memory-based baseline, and the reported experiments show a large character-fidelity advantage over the evaluated alternatives.
Limitations and cautions
- The benchmark is built from narrative-media-derived episodes, so future work could test whether its schedules capture all forms of consistency needed in interactive or documentary video.
- EntityMem depends on verified entity references before generation, which makes reference acquisition an important practical consideration for future systems.
- The benchmark emphasizes characters, objects, and locations; motion style, cinematography, and causal story continuity offer complementary dimensions for future evaluation.
- Even with a strong baseline, the observed degradation at long recurrence distances shows that robust long-form video generation remains an open problem, making the benchmark a useful diagnostic tool.
How to read this result
EntityBench is best read as a carefully structured testbed for a central long-video-generation failure mode, with an interpretable evaluation design and a practical memory baseline that clarifies where current methods still struggle.