PaperlayerView exact revision on arXiv

Source-linked AI summary

ViMax: Agentic Video Generation

Lingxuan Huang, Sizhe He, Hengji Zhou, Liqiang Nie, Lianghao Xia, Chao Huang

arXiv:2606.07649v2cs.CVcs.AI

TL;DR

Current video methods struggle to provide narrative planning and visual consistency across long-form, multi-scene content. ViMax coordinates specialized agents using retrieval-grounded hierarchical planning and graph-based visual dependency mechanisms. It reports improved consistency and narrative-planning metrics over baselines, while noting limits in its models and evaluation.

  • Problem

    Long-form video generation requires narrative planning and visual consistency that current short-clip methods lack, including continuity of characters and environments across scenes.

    Source: §1 · p. 1, §2.2 · p. 4, §2.3 · p. 5

  • Method

    ViMax coordinates specialized production agents and combines hierarchical story planning with retrieval-augmented context and graph-based visual dependency tracking.

    Source: §1 · p. 2, §2.2.2 · p. 5, §2.3.1 · p. 5

  • Results

    Overall GC rises from 0.5272 (HoloCine) to 0.5571, while narrative-planning Avg. improves from 3.827 to 4.107 and NF from 3.08 to 3.52 over AniMaker.

    Source: §3.2 · p. 6

  • Takeaways & Limitations

    ViMax-Bench targets controlled cross-shot consistency challenges, and the framework is presented as a step toward scalable creation of coherent extended narratives.

    Source: §5 · p. 8

    ViMax depends on proprietary foundation models, making performance sensitive to model updates; its automatic evaluation also struggles with fine-grained narrative intent and complex cases.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Long-form video generation requires systematic narrative planning and visual consistency that current short-clip methods cannot provide. Existing methods generate isolated sequences without narrative structure and lack mechanisms for maintaining character and environmental consistency across scenes. We present ViMax, an agentic video generation framework that addresses video creation through coordinated multi-agent collaboration where specialized components negotiate narrative decisions, visual continuity, and production quality. Our framework employs a hierarchical narrative engine with retrieval-augmented generation for global story coherence and a dependency-aware visual consistency mechanism that tracks character and environmental states across temporal boundaries, while VLM-guided agents continuously monitor and refine both narrative coherence and visual fidelity. The framework enables coordinated agent collaboration to generate extended narrative content. This maintains both storytelling integrity and visual coherence across multi-scene timelines.

1 Introduction

Current video generators produce short, isolated clips and struggle to preserve narrative structure and visual consistency across scenes. ViMax coordinates specialized agents for long-form production, combining hierarchical, retrieval-grounded story planning with graph-based visual consistency mechanisms.

1 Introduction

  • Current methods produce short clips without narrative structure, while character appearances and environmental elements can change unpredictably across scenes.Source: §1 · p. 1
  • Long-form generation requires object-level consistency across shots and coordination of character arcs, plot development, and thematic coherence across extended narratives.Source: §1 · pp. 1–2
  • ViMax coordinates agents for screenwriting, shot planning, character styling, video generation, and VLM-based quality control to automate long-form production from conceptual ideas.The framework addresses production from concept through rendered content.Source: §1 · p. 2
  • Hierarchical story decomposition with retrieval-augmented generation grounds local planning in global context to maintain thematic coherence and long-range dependencies.Source: §1 · p. 2
  • Graph-based dependency tracking supports object consistency across shots, while transition video generation maintains spatial coherence across camera angles.Source: §1 · p. 2

2 Methodology

ViMax coordinates specialized agents to turn ideas into long-form videos, combining structured narrative planning with visual-consistency mechanisms. Its workflow organizes planning hierarchically, enriches local story units with global context, and coordinates shot generation and quality assessment.

2.1 Long-Form Video Generation Workflow

  • ViMax coordinates directorial vision, screenwriting, shot planning, character styling, video generation, and quality assessment in a multi-agent workflow.The methodology extends from user ideas to rendered content, with Figure 2 depicting the overall architecture.Source: §2.1 · p. 2, §2 · p. 2, Figure 2 · p. 3

2.1.1 From Ideas to Videos

  • User input becomes a structured screenplay, which is expanded into shot specifications before key images and per-shot videos are generated.Shot specifications include camera angles, movements, lighting, character positioning, and duration; reference images support intermediate visual control.Source: §2.1.1 · pp. 3–4

2.1.2 VLM-based Quality Control

  • A VLM judge scores k parallel candidates against shot requirements, and the highest-scoring candidate advances through the pipeline.Scoring considers visual fidelity, narrative consistency, and adherence to shot specifications.Source: §2.1.2 · p. 4

2.2.1 Hierarchical Story Decomposition

  • ViMax recursively decomposes screenplays into major events, scenes, and shot specifications to make long-form planning manageable.This hierarchy scopes planning tasks while retaining dependencies among narrative units.Source: §2.2.1 · p. 4

2.2.2 RAG-based Global Context Awareness

  • Retrieval-augmented generation indexes the source story and retrieves relevant context to enrich each narrative node’s local description.The approach addresses context lost during decomposition and grounds local decisions in broader story material.Source: §2.2.2 · pp. 4–5

2.3 Visual Consistency Across Shots

  • Independent shot generation can cause character, environment, object, and spatial inconsistencies across long sequences.ViMax introduces complementary visual-coherence mechanisms while seeking to maintain generation efficiency.Source: §2.3 · p. 5

2.3.1 Graph-based Visual Dependency

  • A camera dependency graph captures visual relationships among shots, and dependent cameras use visual anchors propagated from prerequisite nodes.Independent cameras can be generated in parallel; the graph is constructed from shot descriptions and camera assignments.Source: §2.3.1 · p. 5

2.3.2 Spatial Coherence via Transition Videos

  • Transition videos connect camera viewpoints within a scene and serve as spatial anchors for consistent geometry across shots.The method uses transitions between viewpoints rather than generating each independently.Source: §2.3.2 · pp. 5–6

3 Evaluation

Across the reported settings, ViMax leads the consistency and narrative-planning comparisons; the evaluation also examines human preferences, component ablations, efficiency, backbone choice, and quality-control settings.

3.1 Experimental Settings

  • The evaluation uses ViMax-Bench for multi-shot generation and NarrativeQA novels for narrative planning, comparing ViMax with video-generation and planning baselines.The six research questions cover overall performance, human preference, ablations, efficiency, LLM backbone, and hyperparameter sensitivity.Source: §3.1 · p. 6, §3 · p. 6

3.2 Overall Performance Comparison (RQ1)

  • ViMax-Gemini-3-Pro-Veo 3.1 reaches Overall GC 0.5571, versus 0.5272 for HoloCine, 0.4716 for MovieAgent, and 0.5167 for AniMaker, and improves narrative-planning Avg. from 3.827 to 4.107 and NF from 3.08 to 3.52 over AniMaker.The paper reports leads across all nine multi-shot settings and every narrative-planning metric among non-ablated baselines.Source: §3.2 · p. 6, Table 1 · p. 7, Table 3 · p. 7

3.3 Human Evaluation (RQ2)

  • On 35 ViMax-Bench cases, the pairwise human study compares ViMax with five baselines across story groups using cross-scene consistency, semantic following, and aesthetic quality.The supplied passage describes the protocol but gives no win-rate results.Source: §3.3 · p. 6, Table 2 · p. 7

3.4 Ablation Study (RQ3)

  • Removing StoryboardArtist lowers VSU from 4.94 to 3.10 and Avg. from 4.107 to 3.658; removing the Event-Aware Process Chain lowers NF from 3.52 to 2.34, while removing RAG lowers NF to 3.14.The ablations associate these modules with visual articulation, global event structure, and source grounding, respectively; the parallel variant has the largest degradation, lowering Overall GC by 8.7%.Source: §3.4 · p. 6, Figure 3 · p. 7

3.5 Shots Efficiency Study (RQ4)

  • Full ViMax achieves Overall GC 0.5571 in 405s, compared with 0.5085 in 310s for parallel execution and 0.5511 in 835s for serial execution.The system executes independent graph branches concurrently, yielding the reported best consistency–latency balance.Source: §3.5 · p. 7

3.6 Impact of LLM Backbone (RQ5)

  • With the ViMax workflow fixed, GPT-5.4 provides the best trade-off across visual consistency, narrative coherence, and source faithfulness.Claude-Sonnet-4.6 leads on several narrative-quality dimensions, while Gemini-3-Flash ranks second overall.Source: §3.6 · p. 7, Table 4 · p. 8

3.7 Hyperparameter Study (RQ6)

  • Best-of-2 yields Overall GC 0.5571 at $17.500 and 405s per task, while Best-of-3 and Best-of-4 cost more and yield lower GC.The study supports k = 2 as the default operating point; the authors report that additional plausible candidates can increase selection noise in identity and scene state.Source: §3.7 · p. 8, Figure 5 · p. 8

4 Related Work

Related work addresses long-form narrative planning, temporal consistency, and multi-agent content creation through distinct approaches.

Hierarchical Narrative Planning and Cinematic Logic

Multi-Agent Frameworks for Content Creation

5 Conclusion

ViMax combines coordinated agents with hierarchical narrative planning and graph-based visual consistency for long-form video generation.

5 Conclusion

  • ViMax coordinates multi-agent workflows and combines hierarchical story decomposition with retrieval-augmented generation, graph-based dependency tracking, and transition video generation.The dependency graph also exposes shot-level parallelism, and the paper introduces ViMax-Bench for controlled cross-shot consistency challenges.Source: §5 · p. 8
  • The conclusion frames ViMax as a step toward agentic filmmaking, with richer multimodal generation, long-horizon memory, and interactive co-creation as future additions.These are presented as directions for future work, not as capabilities established by the reported system.Source: §5 · p. 8

Limitations

ViMax has limitations in model dependencies, evaluation, generation reliability, and workflow scope.

Limitations

  • ViMax depends on proprietary reasoning, keyframe and video generation, and VLM quality-control models, making performance sensitive to model updates.Its automatic similarity metrics scale well but struggle with fine-grained narrative intent and complex cases; ViMax-Bench also targets controlled challenges at modest scale.Source: Limitations · p. 9
  • The keyframe generator falters in crowded, contact-rich scenes, while video generation may violate keyframe semantics and errors can cascade downstream.Source: Limitations · p. 9
  • The workflow does not yet address audio, dialogue synchronization, shot-level pacing, or interactive revision.Source: Limitations · p. 9
  • Future harnesses may focus on long-horizon visual memory management as commercial models show encouraging short-horizon multi-shot consistency.The authors note that base models may handle some capabilities more strongly natively.Source: Limitations · p. 9

Ethical Considerations

ViMax's stronger cross-shot consistency presents misuse risks, and its safety measures and evaluators have limits.

Ethical Considerations

  • Stronger cross-shot consistency raises risks of impersonation, fabricated events, and imitation of copyrighted characters.Source: Ethical Considerations · p. 9
  • Deployment should require consent, provenance or watermarking, and human review for sensitive cases.Source: Ethical Considerations · p. 9
  • ViMax-Bench safety constraints mitigate benchmark-level risks but not broader misuse, and LLM- and VLM-based evaluators may inherit biases from their models or film sources.The resulting evaluation signals are scalable but imperfect.Source: Ethical Considerations · p. 9

A.1 ViMax-Bench Distribution

Table 5 summarizes ViMax-Bench stories by consistency type and narrative length; the passage specifies the Medium and Long split sizes and shot ranges.

A.1 ViMax-Bench Distribution

  • The Medium split contains 12 stories, each with 2 scenes and 8–10 shots.Source: §A.1 · p. 11
  • The Long split contains 23 stories with 3–4 scenes and 12–16 shots.Source: §A.1 · p. 11
  • Table 5 summarizes ViMax-Bench story composition across consistency type and narrative length.Source: Table 5 · p. 11, §A.1 · p. 11

A.2 Data and Evaluation Protocols

The protocols evaluate controlled multi-shot narrative-to-video challenges using automatic consistency metrics, human comparisons, and narrative-planning scores. Baseline storyboard scoring uses matched 20-storyboard batches.

A.2 Data and Evaluation Protocols

  • ViMax-Bench contains 35 GPT-5.2-generated story specifications spanning 2–4 scenes and 8–16 shots, designed for hard-cut multi-shot generation and controlled consistency challenges.Stories use storyboard panels with first-frame descriptions and motion prompts, cinematic shot designs, recurring characters, and viewpoint shifts; challenge types cover character persistence, background persistence, and multi-person interaction.Source: §A.2 · p. 11
  • Medium stories have 2 scenes and 8–10 shots, while Long stories have 3–4 scenes and 12–16 shots, jointly varying temporal extent and scene complexity.Source: §A.2 · p. 11
  • Automatic consistency metrics encode each shot with ViCLIP features and average pairwise cosine similarity within scenes (IC), across scenes (CC), or across all shots (GC).Higher values indicate greater feature stability; scenario labels define story challenges rather than manually assigned scores.Source: §A.2 · p. 11
  • Human evaluation compares ViMax with five baselines across all 35 stories, using 525 judgments from 15 evaluators on cross-scene consistency, semantic following, and aesthetic quality.Each anonymous pair is judged independently by three evaluators; the study assesses semantic execution and presentation alongside visual continuity.Source: §A.2 · p. 11
  • Novel2Video planning is scored by GPT-5.4 on five storyboard-quality dimensions and narrative faithfulness, with Avg. as their six-dimension mean.Runs use novels sampled from Project Gutenberg's NarrativeQA portion, and the judge applies a 1–5 rubric.Source: §A.2 · p. 12
  • To match the scoring unit, baselines are evaluated on their first 20 sub-storyboards, while ViMax's first 100 are split into five groups of 20 and averaged.This preserves per-batch comparability while reflecting ViMax's long-horizon planning capability.Source: §A.2 · p. 12

A.3 Baseline Details

The section describes ViMax’s baseline comparisons and implementation, then presents analyses and case studies of its visual consistency and narrative planning.

A.3 Baseline Details

  • ViMax uses seven video-generation baselines across four method categories, with all receiving the same 35 stories and shot requirements.Three narrative-planning baselines are evaluated separately under a matched protocol.Source: §A.3 · p. 12

A.4 Implementation Details

  • The shared ViMax workflow uses Gemini models for planning and keyframes, Wan2.2 and Veo 3.1 for video, and VLM selection among generated candidates.Narrative-planning evaluation adds RAG-enhanced context grounding and fixes the compared methods’ LLM backbone to GPT-5.4.Source: §A.4 · p. 13

A.5 Additional Component and Cost Analyses

  • The cost, quality, and latency analysis covers 35 stories, while a four-shot ablation makes component effects on identity, composition, object state, and layout inspectable.The default configuration uses Gemini-3-Flash, Gemini-3-Pro-Image, and Veo 3.1.Source: §A.5 · p. 13, Figure 6 · p. 14, Figure 7 · p. 14

A.6 Case Study

  • ViMax’s case studies show long-horizon character and scene consistency across generated keyframes.Representative segments are shown in Figures 8–10.Source: §A.6 · p. 13

Visual Consistency and Identity Preservation

  • ViMax preserves protagonist identity across noncontinuous shots, camera changes, and four life stages spanning roughly two decades.The examples also show a golden retriever aging from puppy to senior dog.Source: Visual Consistency and Identity… · p. 13
  • Transition video generation acts as a spatial anchor between non-adjacent shots, helping prevent contradictory layouts as the camera repositions.The examples include reverse angles in a police interview and viewpoint changes in other sequences.Source: Visual Consistency and Identity… · p. 14

Narrative Logic and Pacing

Loading 2606.07649v2…