Research map / Video / Video generation
EntityBench: Towards Entity-Consistent Long-Range Multi-Shot Video Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces EntityBench, a 140-episode (2,491-shot) multi-shot video benchmark with per-shot entity schedules and a 51-metric three-pillar evaluation, plus EntityMem, a memory-augmented generation system with per-entity reference banks that improves character fidelity over StoryMem, HoloCine and CineTrans.
Problem
Maintaining consistent characters, objects, and locations across long multi-shot video sequences lacks a standardized benchmark and evaluation; existing benchmarks have few episodes, short shot counts, limited entity types, and simple consistency metrics.
Approach
Derive multi-shot scripts from real narrative media with explicit per-shot entity schedules over three entity types, evaluate cross-shot consistency with embeddings plus LLM per-criterion judgment gated by intra-shot fidelity, and store verified per-entity visual references in a persistent memory bank before generation.
Key results
- EntityMem achieves character presence 0.967 vs 0.882 for the next-best baseline (HoloCine) and face_fidelity 0.740 vs 0.452 for StoryMem.
- On LLM-judged cross-shot face accuracy, EntityMem reaches 0.406 vs 0.226 for StoryMem, while StoryMem leads on DINOv2 cs_face (0.792 vs 0.737).
- Paired effect size for character fidelity vs StoryMem averages Cohen's d = +1.71 across five intra-shot character metrics.
Surprising findings
- EntityMem trails StoryMem on object fidelity and DINOv2 cross-shot embedding similarity, showing embedding metrics and LLM identity judgments can disagree.
- Holistic multi-shot methods achieve higher VBench visual quality (CineTrans wins imaging and aesthetic quality) but much lower character fidelity.
Limitations
- Object consistency is worse than StoryMem's, attributed to incompatibility of entity conditions with keyframe-finetuned weights.
- Evaluation relies on a single proprietary LLM judge (gemini-2.5-pro) and gated metrics that depend on a fidelity threshold.
Experimental setup
Datasets: real narrative media clips (100K production clips, 831 episodes filtered to 606)
Benchmarks: EntityBench
Metrics: VBench dimensions presence per-entity fidelity action fidelity DINOv2 cosine similarity LLM pairwise judgment accuracy Cohen's d
Baselines: StoryMem HoloCine CineTrans
Builds on: StoryMem VBench GroundingDINO
Closest papers by meaning
- MovieBench: A Hierarchical Movie Level Dataset for Long Video Generation (CVPR 2025 · similarity 0.91)
- Advancing Narrative Long Video Generation via Training-Free Identity-Aware Memory (NeurIPS 2026 · similarity 0.89)
- Storybooth: Training-Free Multi-Subject Consistency for Improved Visual Storytelling (ICLR 2025 · similarity 0.88)
- CineOrchestra: Unified Entity-Centric Conditioning for Cinematic Video Generation (NeurIPS 2026 · similarity 0.88)
- Storynizor: Consistent Story Generation via Inter-Frame Synchronized and Shuffled ID Injection (AAAI 2025 · similarity 0.87)
- StoryDiffusion: Consistent Self-Attention for Long-Range Image and Video Generation (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.