PersonaShot — Narrative Continuity in Multi-Shot Video Generation

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.

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. We distill reasoning from a large multimodal teacher (Qwen3.5-397B) into lightweight specialist evaluators (Qwen3.5-4B + LoRA), each grounded in the evidence required by its metric, and align them with expert human judgments.

Our evaluation reveals distinct capability profiles across state-of-the-art models and a clear gap between perceptual quality and cross-shot narrative continuity.

1,000+

Multi-Shot Segments

5,000+

Annotated Shots

Fine-Grained Metrics

Narrative Dimensions

5.3

Avg Shots / Segment

ρ=.73

Human Alignment

Three Narrative Dimensions

Causal Physical Continuity

Spatial Layout, Object State Persistence, Geometric Scale — do character positions, held objects, and proportions remain causally consistent across editing cuts?

Affective Dynamics

Expression Naturalness, Facial-Action Temporal Coherence, Emotion-Narrative Alignment, Emotional Arc — do affective states evolve coherently without abrupt resets?

Cinematic Grammar

Shot Transition, Transition Rhythm, 180° Rule, Eyeline Match, Directorial Narrative Sequencing — does the sequence follow film-language conventions?

Overview

Top: Qualitative examples illustrating three core dimensions. Bottom: Quantitative comparison revealing distinct capability profiles of state-of-the-art models.

Data Pipeline

Three-Stage Processing

Stage 1: Sequence Segmentation & Character Continuity (TransNetV2, 2–15 shots/segment). Stage 2: Face Visibility & Quality (RetinaFace + CLIP-IQA). Stage 3: Cross-Shot Relation Selection (Qwen-based multimodal reasoning).

Evaluation Framework

PersonaShot extends conventional quality assessment with three specialist dimensions. Lightweight evaluators (Qwen3.5-4B + LoRA) are distilled from Qwen3.5-397B via criterion-specific evidence.

Specialist Evaluator Design

Physical Continuity: Teacher-guided distillation with strict input isolation — annotations used during training but excluded at inference.

Affective Dynamics: Frame-level Action Unit signals from OpenFace enable AU-aware evaluation based on the Facial Action Coding System (FACS).

Cinematic Grammar: Two-stage pipeline: CineCap extracts professional film-language descriptions, then a reasoner evaluates via TransNetV2 shot boundaries.

Leaderboard

Click any column header to sort. Bold = best, underlined = second best.

Method	Aggregate	Visual Quality and Consistency				Causal Physical Continuity			Affective Dynamics				Cinematic Grammar
Method	Aggr.	VF	TVA	ID	SS	SLC	OSP	GSC	EN	FATC	ENA	EAC	ST	TRA	180R	EM

bold best · underline second. VF=Visual Fidelity, TVA=Text-Video Alignment, ID=Identity, SS=Scene/Style, SLC=Spatial Layout, OSP=Object State Persistence, GSC=Geometric Scale, EN=Expression Naturalness, FATC=Facial-Action Temporal Coherence, ENA=Emotion-Narrative Alignment, EAC=Emotional Arc, ST=Shot Transition, TRA=Transition Rhythm, 180R=180° Rule, EM=Eyeline Match.

Key Findings

1. Perceptual quality ≠ narrative continuity

ShotStream achieves highest visual fidelity (.994 VF) yet ranks 7th overall (.601) due to weak affective and cinematic performance.

2. Narrative reasoning matters more than prompt granularity

Seedance leads overall (.793). Global-prompt systems remain competitive in cinematic grammar and identity.

3. Physical & affective state propagation remain major challenges

Models struggle with object-state resets, unstable facial dynamics, and weak emotional trajectories across cuts.

4. Cinematic grammar is largely under-modeled

Most systems fail at transition appropriateness, 180° rule compliance, and eyeline matching.

5. Stark divide in long-form cinematic control

Seedance (0.865 DNS) vs. open-source LTX-2.3 (0.624) and VGoT (0.535) — long-form narrative planning remains a bottleneck.

Qualitative Failure Cases

Representative failures across three dimensions: object state resets, emotion inconsistencies, eyeline mismatches, and inappropriate transitions. Many failures are visually plausible per-shot but emerge across temporal and relational evaluation.

Human Study & Evaluator Validation

Setup: 10 domain experts, 25 multi-shot sequences, 4 annotators per sequence. Krippendorff’s α = 0.76. All correlations p < 0.01.

Dim.	Criterion	Spearman ρ ↑	Pair Acc. ↑
Physical	Spatial Layout (SLC)	0.68	71.4%
	Object State (OSP)	0.74	76.0%
	Geometric Scale (GSC)	0.66	69.8%
Affective	Expression Naturalness (EN)	0.70	71.8%
	Action Coherence (FATC)	0.68	64.8%
	Narrative Alignment (ENA)	0.71	72.5%
	Arc Coherence (EAC)	0.67	68.2%
Cinematic	Narrative Sequencing (DNS)	0.74	74.0%
	Shot Transition (ST)	0.72	73.1%
	Rhythm Alignment (TRA)	0.69	70.5%
	180° Rule (180R)	0.71	71.9%
	Eyeline Match (EM)	0.65	67.2%

All 12 specialist criteria evaluated against expert MOS. DNS (ρ=0.74) and OSP (ρ=0.74) show strongest human alignment.

Qualitative Comparison

Comparison videos coming soon. An IVEBench-style group carousel will be placed here, showing a source clip followed by every competing method's output with category / edit-prompt labels.

Sample Visualizations

Conclusion & BibTeX

PersonaShot is the first person-centric benchmark for narrative continuity in multi-shot video generation. Through 16 metrics and human-aligned specialist evaluators, it reveals that current generators remain limited in preserving persistent character states, modeling emotional evolution, and following cinematic conventions across shots. We hope PersonaShot serves as both a rigorous testbed and a catalyst for next-generation video foundation models with authentic storytelling capabilities.

@misc{personashot2026,
  title     = {PersonaShot: Benchmarking Person-Centric Narrative Continuity in Multi-Shot Video Generation},
  author    = {Wang, Yuji and Chen, Yuheng and Hu, Teng and Yi, Ran and Hong, Yijia and Feng, Han and Cao, Weijian and Wang, Chengjie and Ma, Lizhuang and Zhang, Jiangning},
  year      = {2026},
  eprint    = {arXiv:XXXX.XXXXX},
  archivePrefix = {arXiv},
  primaryClass = {cs.CV}
}