PersonaShot: Benchmarking Person-Centric Narrative Continuity in Multi-Shot Video Generation

Yuji Wang; Yuheng Chen; Teng Hu; Ran Yi; Yijia Hong; Han Feng; Weijian Cao; Chengjie Wang

cs.CVAug 17, 2026

PersonaShot: Benchmarking Person-Centric Narrative Continuity in Multi-Shot Video Generation

Authors: Yuji Wang, Yuheng Chen, Teng Hu, Ran Yi, Yijia Hong, Han Feng, Weijian Cao, Chengjie Wang,

Organizations: Shanghai Jiao Tong University · Tencent Youtu Lab · Zhejiang University

Abstract

Video generation is rapidly evolving from single-shot clips to multi-shot narratives, where the human character serves as the core narrative anchor. However, existing benchmarks mainly assess character appearance or individual-shot quality, without measuring whether physical and emotional states remain coherent across cuts. They also rarely provide criterion-specific evaluation methods, although physical continuity, facial dynamics, and cinematic relations require different visual, temporal, and relational evidence. To address these limitations, we introduce PersonaShot, the first person-centric benchmark for narrative continuity in multi-shot video generation. PersonaShot contains approximately 1,000 multi-shot segments and 16 metrics spanning physical continuity, affective dynamics, and cinematic grammar. 1) Narrative Continuity Benchmark: We evaluate character coherence across three temporal levels: within-shot states, cross-shot transitions, and sequence-level trajectories. 2) Human-Aligned Specialist Evaluators: We distill reasoning from a large multimodal teacher into lightweight criterion-specific evaluators, each grounded in the visual, temporal, or relational evidence required by its metric, and align them with expert human judgments. 3) Systematic Evaluation and Insights: Our evaluation reveals distinct capability profiles across state-of-the-art models and a clear gap between perceptual quality and cross-shot narrative continuity. Even visually compelling videos frequently exhibit physical-state resets, abrupt affective shifts, and broken cinematic relations across shots. Human studies further demonstrate strong agreement between our evaluators and expert judgments.

Figures & tables

Figure 1: Overview of PersonaShot’s person-centric evaluation. Top: Qualitative examples illustrating our three core dimensions: physical continuity, affective dynamics, and cinematic grammar. Bottom: Quantitative comparison across fine-grained metrics for specific state-of-the-arts, revealing their distinct capability profiles.

Benchmark	Input Modalities				Evaluation Capabilities
	Ref. Image	Ref. Audio	First- Frame	Multi- Shot	Person- Centric	Physics		Emotion		Cinematic Grammar
	Ref. Image	Ref. Audio	First- Frame	Multi- Shot	Person- Centric	Intra	Cross	Coarse	AU-Aware	Cinematic Grammar
VBench [ 11 ]	$\times$	$\times$	$\times$	$\times$	$\times$	$\times$	$\times$	$\times$	$\times$	$\times$
HumanVBench [ 39 ]	$\checkmark$	$\times$	$\times$	$\times$	$\checkmark$	$\times$	$\times$	$\checkmark$	$\times$	$\times$
HVEval [ 32 ]	$\times$	$\times$	$\times$	$\times$	$\checkmark$	$\times$	$\times$	$\checkmark$	$\times$	$\times$
VideoPhy-2 [ 3 ]	$\checkmark$	$\times$	$\times$	$\times$	$\times$	$\checkmark$	$\times$	$\times$	$\times$	$\times$

Table 1: Comparison of PersonaShot with existing video generation benchmarks. First-Frame denotes shot-level image conditioning; Intra/Cross denote within-shot plausibility and cross-shot continuity; AU-Aware uses facial Action Units, while Cinematic Grammar covers the 180-degree rule, eyeline matching, and transition logic.

Figure 2: Overview of the PersonaShot benchmark. Top-Left: Our three-stage data processing pipeline filters multi-shot segments via sequence continuity, face visibility, and cross-shot relations. Top-Right: A curated sample featuring hierarchical text annotations and multi-modal conditioning inputs. Bottom: Key benchmark statistics.

Figure 3: Overview of PersonaShot evaluation framework. Given generated multi-shot videos and prompts, PersonaShot extends conventional quality assessment with three specialist dimensions for narrative-level evaluation: causal physical continuity, affective dynamics, and cinematic grammar.

Method	Visual Quality and Consistency				Causal Physical Continuity			Affective Dynamics				Cinematic Grammar				Aggregate
Method	VF	TVA	ID	SS	SLC	OSP	GSC	EN	FATC	ENA	EAC	ST	TRA	180R	EM	Aggregate
Shot-Level Prompt Conditioning
Seedance	0.981	0.752	0.673	0.762	0.778	0.928	0.947	0.907	0.562	0.778	0.617	0.837	0.658	0.818	0.813	0.793
HoloCine	0.853	0.797	0.582	0.718	0.647	0.862	0.893	0.792	0.384	0.677	0.343	0.723	0.517	0.712	0.687	0.687
STAGE	0.827	0.703	0.507	0.752	0.623	0.818	0.852	0.718	0.453	0.628	0.443	0.677	0.503	0.667	0.642	0.661
LTX-2.3	0.892	0.718	0.572	0.733	0.677	0.832	0.857	0.758	0.427	0.583	0.403	0.683	0.512	0.683	0.648	0.673

Table 2: Fine-grained evaluation results on PersonaShot under shot-level and global-prompt conditioning. Higher is better. The aggregate score equally averages the four dimension scores, with metrics averaged within each dimension. Within each protocol, the best and second-best results are highlighted in bold and underlined , respectively.

Figure 4: Qualitative failure cases revealed by PersonaShot . The examples illustrate representative failures in causal physical continuity, affective dynamics, and cinematic grammar, including object state persistence failures, spatial layout discontinuities, emotion evolution inconsistencies, facial expression issues, eyeline matching failures, and inappropriate shot transitions.

Dimension	Criterion	$\boldsymbol{\rho}\uparrow$	Pair. $\uparrow$
Physical	Spatial Layout (SL)	0.68	71.4%
	Object State (OSP)	0.74	76.0%
	Aggregate	0.75	75.2%
Affective	Emotion Alignment (ENA)	0.71	72.5%
	Facial Dynamics (FATC)	0.68	64.8%
	Aggregate	0.69	70.3%

Table 3: Human alignment of PersonaShot evaluators. We report Spearman correlation (

\rho

) with human Mean Opinion Scores (

N=25

) and pairwise agreement (Pair.) for representative criteria. Full results are in the Appendix E.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Hyperparameter	Qwen3-4B Evaluator	Qwen3-8B Evaluator
Base Model Backbone	Qwen/Qwen3-4B	Qwen/Qwen3-8B
Training / Validation Samples	4,041 / 446 (MERR dataset)
Target Emotion Categories	9 classes
LoRA Rank ( $r$ )	16	32
LoRA Alpha ( $\alpha$ )	32	64
Trainable Parameters	33M (0.81%)	66M ( $\sim$ 0.80%)

Appendix

Table 4: LoRA Fine-Tuning Configurations for Specialist Evaluators. Training setup and hyperparameters for adapting Qwen3-4B and Qwen3-8B on the 4,041-sample MERR training set.

Dimension	Metric	Definition	Evaluation Strategy
Visual Quality and Consistency
Visual Quality and Consistency	Visual Fidelity (VF)	Measures perceptual quality, artifacts, distortions, and composition.	DOVER aesthetic/technical assessment and MUSIQ frame-level quality evaluation.
	Text-Video Alignment (TVA)	Measures whether generated content follows textual instructions, including entities, actions, and scenes.	VQAScore semantic matching combined with GroundingDINO-based entity grounding.
	Identity Consistency (ID)	Measures whether recurring characters preserve visual identity across shot transitions.	Character localization and tracking followed by cross-shot identity embedding similarity.
	Scene & Style Consistency (SS)	Measures background, lighting, color, and visual style stability across shots.	Background similarity, style embedding comparison, and multimodal visual consistency judgments.
	Reference Fidelity (RF)	Measures preservation of identity from reference images or audio conditions.	DINOv2 and ArcFace similarities for visual identity preservation.

Appendix

Table 5: Detailed evaluation metrics of PersonaShot (Part I). The table summarizes the foundational visual metrics and causal physical continuity metrics.

Dimension	Metric	Definition	Evaluation Strategy
Affective Dynamics
Affective Dynamics	Expression Naturalness (EN)	Measures facial realism and generation artifacts in expressions.	Multimodal expression assessment combined with facial Action Unit (AU) analysis.
	Facial-Action Temporal Coherence (FATC)	Measures temporal stability of facial actions and detects jitter or unnatural static expressions.	Frame-level AU trajectories are analyzed through temporal variation statistics.
	Emotion-Narrative Alignment (ENA)	Measures whether emotional category and intensity match the narrative context.	Narrative affect extracted from annotations is compared with AU-aware facial affect estimation.
	Emotional Arc Coherence (EAC)	Measures whether emotional states evolve coherently throughout the sequence.	Affective trajectories are constructed across shots and compared with expected narrative evolution.
Cinematic Grammar

Appendix

Table 6: Detailed evaluation metrics of PersonaShot (Part II). The table summarizes affective dynamics and cinematic grammar metrics.

Rank	Model / System	Composite ( $\uparrow$ )	D1: Visual ( $\uparrow$ )	D2: Physical ( $\uparrow$ )	D3: Emotion ( $\uparrow$ )	D4: Grammar ( $\uparrow$ )
Shot-Level Prompt Conditioning
1	Seedance (Structured)	0.730	0.751	0.824	0.630	0.716
2	HoloCine (Storyboard)	0.707	0.764	0.823	0.594	0.646
3	STAGE + Wan2.2	0.650	0.673	0.774	0.546	0.609
4	LongLive	0.623	0.642	0.755	0.453	0.642
5	MultiShotMaster	0.617	0.639	0.755	0.546	0.530

Appendix

Table 7: Complete 12-Model Dimension Leaderboard on PersonaShot. We report the aggregate Composite score alongside four macro-dimensions across 100 standardized test sequences per model. Models are categorized into Shot-Level Prompt Conditioning and Global-Prompt Conditioning . Equal dimension weighting (

\alpha=0.25

) is applied. Within each protocol, the highest score in each column is highlighted in bold .

Model / System	EPA ( $\uparrow$ )	Hard Acc ( $\uparrow$ )	Soft Acc ( $\uparrow$ )	Traj Corr ( $\uparrow$ )
Seedance (Structured)	0.802	0.824	0.869	0.648
STAGE + Wan2.2	0.797	0.816	0.867	0.657
CineTrans	0.797	0.820	0.870	0.646
MultiShotMaster	0.796	0.815	0.864	0.651
ShotStream	0.796	0.814	0.863	0.651
HoloCine	0.793	0.815	0.864	0.637

Appendix

Table 8: Emotion-Prompt Alignment (EPA) Results. We report the composite EPA score alongside exact label match (Hard Acc), valence-arousal similarity (Soft Acc), and emotional arc trajectory correlation (Traj Corr) for shot-level prompt conditioning systems. Best performance is highlighted in bold .

Dimension	Per-Shot Avg ( $\uparrow$ )	Cross-Shot Avg ( $\uparrow$ )	Absolute Drop ( $\Delta$ )
D1: Visual Quality	0.780	0.580	-0.200 (-25.6%)
D3: Emotional Expression	0.700	0.350	-0.350 (-50.0%)

Appendix

Table 9: Per-Shot vs. Cross-Shot Performance Degradation. Averaged scores across all evaluated models show substantial drops when transitioning from local shot fidelity to cross-shot relational continuity.

Prompt Condition	Composite ( $\uparrow$ )	D1: Visual ( $\uparrow$ )	D2: Physical ( $\uparrow$ )	D3: Emotion ( $\uparrow$ )	D4: Grammar ( $\uparrow$ )
LTX-2 Full Prompt ( $\sim$ 1200 tokens)	0.705	0.722	0.786	0.482	0.744
LTX-2 Brief Prompt ( $\sim$ 600 tokens)	0.464	0.434	0.409	0.269	0.634
Absolute Degradation ( $\Delta$ )	-0.241	-0.288	-0.378	-0.212	-0.110
Relative Performance Drop	-34.2%	-39.9%	-48.0%	-44.1%	-14.8%

Appendix

Table 10: Impact of Prompt Length on LTX-2 Performance. Halving storyboard prompt length causes severe degradation in physical and visual coherence while leaving cinematic grammar relatively resilient.

Figure 5: Qualitative Multi-Model Comparison on a 3-Shot Emotional Dialogue Scene. Comparison of nine video generation systems given a hierarchical prompt describing an intense emotional interaction between a young man and a middle-aged woman. While Seedance and HoloCine maintain character identities and emotional distress across cuts, baselines exhibit severe failure modes: ShotStream replaces the character with an irrelevant vase, LTX-2.3 suffers from monochrome style drift, and EchoShot experiences sharp identity shifts across shot boundaries.

Figure 6: Qualitative Multi-Model Comparison on a 4-Shot Comical Action Sequence. Comparison across nine systems generating a multi-shot narrative of a man driving a car, peeling a stamp, licking it, and smiling. Seedance and HoloCine accurately execute the fine-grained object interaction and facial expressions while preserving character identity. In contrast, LongLive hallucinates a child identity in Shot 3, ShotStream generates a bedroom alarm clock instead of the car interior, and STAGE struggles with stamp-envelope object persistence.

Figure 7: Coefficient of Variation (CV %) across PersonaShot Sub-Metrics. Higher CV indicates stronger discriminative power across evaluated systems. Emotion Narrative Alignment ( $25.3\%$ ) exhibits the highest sensitivity across all models.

Figure 8: Internal Correlation Heatmaps (Pearson’s $\boldsymbol{r}$ ) for Specialist Dimensions. From left to right: Physical Consistency ( D2 ), Emotional Expression ( D3 ), and Film Grammar ( D4 ). Low correlations across multiple sub-metrics confirm the orthogonality and non-redundancy of our benchmark criteria.

Dimension	$\boldsymbol{N}$	$\boldsymbol{\rho}\uparrow$	Pair Acc. $\uparrow$
Causal Physical Continuity	25	0.75	75.2%
Affective Dynamics	25	0.69	70.3%
Cinematic Grammar	25	0.73	72.8%
Overall PersonaShot	25	0.78	76.8%

Appendix

Table 11: Dimension-level alignment with human MOS. All correlations are statistically significant at

p<0.01

Dim.	Criterion	$\boldsymbol{\rho}\uparrow$	Pair Acc. $\uparrow$
Phy.	Spatial Layout (SLC)	0.68	71.4%
	Object State (OSP)	0.74	76.0%
	Geometric Scale (GSC)	0.66	69.8%
Aff.	Expr. Naturalness (EN)	0.70	71.8%
	Action Coherence (FATC)	0.68	64.8%
	Narrative Align. (ENA)	0.71	72.5%

Appendix

Table 12: Fine-grained criterion-level human alignment. Comparison of automatic evaluator scores against expert MOS across all 12 specialist criteria. Correlations are computed over valid evidence-gated sequences (

p<0.01

Video Generation

Video generation studies models that create and edit moving images from text, images, audio, and structured controls.

See all 354 papers

MSVBench: Towards Human-Level Evaluation of Multi-Shot Video Generation
Feb 27, 2026Haoyuan Shi, Yunxin Li, Nanhao Deng +5Video Generation Evaluation Multi-Shot Video Generation
Harbin Institute of Technology (Shenzhen) · Alibaba International Digital Commerce
EntityBench: Towards Entity-Consistent Long-Range Multi-Shot Video Generation
May 14, 2026Ruozhen He, Meng Wei, Ziyan Yang +1Long-Horizon Video Generation Memory-Augmented Video Generation
ByteDance · Rice University · ByteDance Seed
CANVAS: Continuity-Aware Narratives via Visual Agentic Storyboarding
Apr 15, 2026Ishani Mondal, Yiwen Song, Mihir Parmar +4Visual Storytelling Narrative Generation
University of Maryland, College Park · Google

Feb 27, 2026J/K move · Enter open · S save

MSVBench: Towards Human-Level Evaluation of Multi-Shot Video Generation

Haoyuan Shi, Yunxin Li, Nanhao Deng, Zhenran Xu+4

Harbin Institute of Technology (Shenzhen) · Alibaba International Digital Commerce

Video Generation Evaluation Multi-Shot Video Generation Temporal Consistency in Video Generation

Open paper AskPDF arXiv

The evolution of video generation toward complex, multi-shot narratives has exposed a critical deficit in current evaluation methods. Existing benchmarks remain anchored to single-shot paradigms, lacking the comprehensive story assets and cross-shot metrics required to assess long-form coherence and appeal. To bridge this gap, we introduce MSVBench, the first comprehensive benchmark featuring hierarchical scripts and reference images tailored for Multi-Shot Video generation. We propose a hybrid evaluation framework that synergizes the high-level semantic reasoning of Large Multimodal Models (LMMs) with the fine-grained perceptual rigor of domain-specific expert models. Evaluating 20 video generation methods across diverse paradigms, we find that current models--despite strong visual fidelity--primarily behave as visual interpolators rather than true world models. We further validate the reliability of our benchmark by demonstrating a state-of-the-art Spearman's rank correlation of 94.4% with human judgments. Finally, MSVBench extends beyond evaluation by providing a scalable supervisory signal. Fine-tuning a lightweight model on its pipeline-refined reasoning traces yields human-aligned performance comparable to commercial models like Gemini-2.5-Flash.

Abstract

Video Generation

Explore similar work

MSVBench: Towards Human-Level Evaluation of Multi-Shot Video Generation

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

CANVAS: Continuity-Aware Narratives via Visual Agentic Storyboarding