PaperlayerView exact revision on arXiv

Source-linked AI summary

Camera Artist: A Multi-Agent Framework for Cinematic Language Storytelling Video Generation

Haobo Hu, Qi Mao, Yuanhang Li, Libiao Jin

arXiv:2604.09195v1cs.AI

TL;DR

Existing video-generation approaches prioritize local clip quality, and multi-agent systems can lack mechanisms for continuity across adjacent shots and deliberate cinematic language. Camera Artist integrates recursive shot planning and cinematic language injection into an automated filmmaking workflow. The paper reports stronger narrative coherence, shot diversity, and temporal stability than existing baselines.

  • Problem

    Current video models emphasize local visual quality, while multi-agent systems may lack explicit mechanisms for shot-to-shot continuity and cinematic language.

    Source: §I · p. 1

  • Method

    Camera Artist automates narrative video generation with recursive shot planning conditioned on prior shots and cinematic language injection into shot descriptions.

    Source: §I · p. 2, B. Recursive Shots Generation · p. 3

  • Results

    Camera Artist is reported to achieve superior narrative coherence, shot diversity, and temporal stability compared with existing baselines.

    Source: §I · p. 2

  • Takeaways & Limitations

    The paper presents Camera Artist as a framework for cinematic narrative generation that combines narrative coherence with film-level visual expressiveness.

    Source: §IV · p. 6

Abstract

from arXiv · show

We propose Camera Artist, a multi-agent framework that models a real-world filmmaking workflow to generate narrative videos with explicit cinematic language. While recent multi-agent systems have made substantial progress in automating filmmaking workflows from scripts to videos, they often lack explicit mechanisms to structure narrative progression across adjacent shots and deliberate use of cinematic language, resulting in fragmented storytelling and limited filmic quality. To address this, Camera Artist builds upon established agentic pipelines and introduces a dedicated Cinematography Shot Agent, which integrates recursive storyboard generation to strengthen shot-to-shot narrative continuity and cinematic language injection to produce more expressive, film-oriented shot designs. Extensive quantitative and qualitative results demonstrate that our approach consistently outperforms existing baselines in narrative consistency, dynamic expressiveness, and perceived film quality.

I. INTRODUCTION

Current video models prioritize high-fidelity short clips, while multi-agent filmmaking systems can still lack shot-to-shot continuity and deliberate cinematic expression. Camera Artist adds recursive shot planning and cinematic language injection; the paper reports stronger storytelling outcomes than existing baselines.

I. INTRODUCTION

  • Existing text-to-video and image-to-video models prioritize local visual quality over the cinematic reasoning needed to orchestrate multi-stage narratives.The paper frames bridging visually striking fragments and coherent cinematic narratives as a central challenge.Source: §I · p. 1
  • Multi-agent systems can align scripts and visuals yet still produce mechanical scene assembly without deliberate cinematic authorship.The paper asks how such systems can move beyond storytelling sequences to videos that feel like cinema.Source: §I · p. 1
  • The introduction contrasts Camera Artist’s stronger shot-to-shot coherence and richer cinematic expression with fragmented narratives and weak cinematic control in existing multi-agent methods.The comparison is presented in Fig. 1.Source: Figure 1 · p. 1
  • Camera Artist combines recursive shot generation, conditioned on preceding-shot context, with a specialized language model that turns plot points into film-oriented technical descriptions.The paper presents these as mechanisms for narrative continuity and cinematic expression.Source: §I · p. 2
  • The framework automates narrative video generation from script understanding through cinematic shot planning to final rendering.Its stated contributions include explicit recursive shot generation and cinematic language injection.Source: §I · p. 2
  • Camera Artist is reported to outperform existing baselines in narrative coherence, shot diversity, and temporal stability.The introduction describes these as results of extensive experiments.Source: §I · p. 2

II. OUR SOLUTION: CAMERA ARTIST

Camera Artist builds on established multi-agent filmmaking workflows to turn a story outline into an ordered sequence of video clips. Its collaborative agents divide narrative planning, cinematic shot design, and visual rendering across a hierarchical storyboard pipeline.

II. OUR SOLUTION: CAMERA ARTIST

  • Camera Artist transforms a user-provided story outline into a temporally ordered sequence of video clips, targeting shot-level narrative coherence and cinematic expressiveness.The framework builds on established multi-agent filmmaking workflows rather than replacing the agentic paradigm.Source: §II · p. 2

A. Multi-Agent Collaborative System Framework

  • Three agents handle narrative planning, cinematic shot design, and visual rendering across footage-construction and shot-generation stages.The Director Agent builds hierarchical storyboard assets, the Cinematography Shot Agent recursively generates and enriches shot descriptions, and the Video Generation Agent produces and stitches clips.Source: A. Multi-Agent Collaborative System… · p. 2, Figure 3 · p. 3
  • The Director Agent decomposes the outline into script-level resources, scene properties, and visual references, which inform ordered shot descriptions enriched with cinematic attributes.The Video Generation Agent conditions a multi-reference image-to-video model on references and shot descriptions, then concatenates clips into the long-form video.Source: A. Multi-Agent Collaborative System… · p. 3

B. Recursive Shots Generation

Recursive storyboard generation plans shots using the global script, scene context, and preceding-shot information. It distinguishes scene-start, midpoint, and endpoint shots to structure the sequence and its continuity.

B. Recursive Shots Generation

  • Each shot is planned from the global script and prior-shot context, simulating how sequential shots are connected during human writing.The Cinematography Shot Agent generates shots in scene order and determines each shot’s content and type using the scene and prior-shot information.Source: B. Recursive Shots Generation · p. 3
  • The recursive process distinguishes a scene start generated without a previous shot, midpoints conditioned on the previous shot, and a scene endpoint.The endpoint marks the end of recursive shot generation for that scene.Source: B. Recursive Shots Generation · p. 3
  • For a scene, shot generation uses the scene and preceding shot as context and stops when the endpoint shot is reached.The passage describes the shots as generated recursively.Source: B. Recursive Shots Generation · p. 3

C. Cinematic Language Injection

Cinematic Language Injection extends recursive shot generation by refining shots with purposeful camera attributes. At inference, a fine-tuned LLM adds explicit cinematic semantics to recursively generated shot descriptions.

C. Cinematic Language Injection

  • Cinematic Language Injection refines each shot with purposeful camera attributes to express professional cinematic language and visual intention.The module is built upon Recursive Shot Generation.Source: C. Cinematic Language Injection · p. 3
  • Camera Artist produces richer, more expressive cinematic language than prior multi-agent methods for videos with similar shot content.The comparison is qualitative and concerns single-shot content.Source: Figure 4 · p. 4
  • A fine-tuned LLM injects explicit cinematic semantics into each recursively generated shot description during inference.The training description maps ordinary video descriptions and shot-level cinematic annotations to cinematic-enriched descriptions.Source: C. Cinematic Language Injection · pp. 3–4

III. EXPERIMENTS

The experiments use a common agent backbone and video-generation setup, evaluate MoviePrompts and an additional storytelling benchmark, and assess perceptual and semantic dimensions.

A. Experimental Setup

  • All agents use Qwen3-30B-A3B-Instruct, while Qwen3-4B is fine-tuned with LoRA on 580 ShotBench paired samples for cinematic language injection.Fine-tuning runs for 20 epochs; MAGREF generates video clips and Flux creates reference images.Source: A. Experimental Setup · p. 4
  • Evaluation uses MoviePrompts, containing plot descriptions and character profiles from ten professional films, plus eight additional storytelling samples.The additional benchmark follows the same format as MoviePrompts.Source: A. Experimental Setup · p. 4
  • The evaluation covers VBench perceptual dimensions and CLIP-T semantic similarity; Table I compares VBench and CLIP-based semantic consistency.VBench dimensions include subject and background consistency, motion smoothness, dynamic degree, and aesthetic score.Source: A. Experimental Setup · p. 4, Table I · p. 4

B. Comparison with Baseline

Camera Artist is reported to perform strongly across the evaluated metrics and narrative and cinematic dimensions. Qualitative comparisons describe more expressive single shots and more coherent transitions between shots.

B. Comparison with Baseline

  • Camera Artist performs strongly across evaluated metrics, achieves the highest motion dynamics while maintaining high background consistency, and is rated exceptionally well by VLM evaluators across narrative and cinematic dimensions.VGoT [10] has the highest subject consistency, which the authors attribute to near-static videos and its lowest dynamic-degree scores.Source: B. Comparison with Baseline · p. 4, Table II · p. 5
  • In a qualitative comparison of inter-shot narrative coherence, Camera Artist conditions each shot on preceding-shot and scene information, producing shot content narratively coherent in both text and visual realization.Source: Figure 5 · p. 5
  • For the “Elsa senses magical energy” prompt, Camera Artist uses a high-angle wide shot with a smooth zoom-out, while Anim-Director, VGoT [10], and MovieAgent [11] produce less expressive shots.The cited examples include visually similar shots, a fixed mid-to-long shot, and a largely static close-up, respectively.Source: B. Comparison with Baseline · p. 5
  • In the user-study comparison, Camera Artist receives the highest scores across Script Consistency, Camera-Movement Consistency, Video Quality, and Real-Movie Similarity.The figure compares Camera Artist with VGoT [10], Anim-Director [8], and MovieAgent [11].Source: Figure 6 · p. 5

C. User Study

Human evaluation reports the highest aggregate scores for Camera Artist across four subjective dimensions, while ablations associate RSG and CLI with distinct aspects of the results. The conclusion describes the framework as improving narrative coherence and film-level visual expressiveness.

C. User Study

  • Camera Artist scores 4.28 in Script Consistency and 4.12 in Real-Movie Similarity, and has the highest aggregate scores across all four human-evaluation dimensions.The five-point Likert study also assesses Camera-Movement Consistency and Video Quality.Source: C. User Study · p. 5
  • Removing RSG yields the lowest script-consistency performance and weaker inter-shot coherence, while removing CLI drops camera-motion fidelity from 3.55 to 2.83.The CLI ablation produces largely static, descriptive videos; the full model combines coherent narrative with deliberate camera motion, angles, and lighting.Source: C. User Study · p. 6, Figure 7 · p. 6

IV. CONCLUSIONS

  • The conclusion states that integrating recursive storyboard generation and cinematic language injection improves narrative coherence and film-level visual expressiveness beyond conventional clip-centric generation.It also describes Camera Artist as a framework for cinematic narrative generation.Source: §IV · p. 6

APPENDIX

The appendix outlines Camera Artist’s workflow, CLI fine-tuning, and the reasoning prompts used by its agents. It also previews supplementary implementation and qualitative results.

APPENDIX

  • The appendix adds implementation details, evaluation metrics, user-study and baseline information, and further qualitative results.Source: APPENDIX · p. 7

A. Workflow Overview

  • The Director Agent turns a story outline into scene plots, character assets, and reference images; the Cinematography Shot Agent plans contextualized, cinematic shots for video rendering.The Video Generation Agent renders clips that are temporally concatenated into a long-form narrative.Source: A. Workflow Overview · p. 7

B. Cinematic Language LoRA Fine-tuning.

  • CLI fine-tuning uses 580 paired ordinary and cinematic descriptions from ShotBench annotations to adapt Qwen3-4B with LoRA for inference-time cinematic attribute injection.The setup uses rank 8, scaling factor 32, learning rate 1 × 10^-4, and 20 epochs.Source: B. Cinematic Language LoRA Fine-tuning · p. 7

C. Details of the Chain-of-Thought (CoT) Prompts

  • Director Agent prompts guide progressive planning of genre, characters, scene objectives, and scene decomposition, while shot-agent prompts reason over prior shots and current scene intent.The shot agent uses this reasoning for recursive storyboard generation and cinematic decisions rather than one-step shot output.Source: C. Details of the Chain-of-Thought… · p. 7

D. Details of Evaluation Details

The section describes the evaluation measures and procedures used to assess generated videos. It covers automatic metrics, VLM scoring, and a randomized human-rating study.

D. Details of Evaluation Details

  • Automatic evaluation combines VBench perceptual dimensions, CLIP-T script similarity, and frame-level CLIP image-text alignment.VBench dimensions include subject and background consistency, motion smoothness, dynamic degree, and aesthetic score.Source: D. Details of Evaluation Details · p. 7
  • VLM evaluators score script consistency, camera-movement consistency, video quality, and real-movie similarity on a 1–5 scale using task-specific prompts.Videos are uniformly sampled into 8–12 keyframes, provided with the script or camera-motion plan, and scored with a rubric and JSON output format.Source: D. Details of Evaluation Details · p. 7
  • The workflow visualization shows story-outline planning by the Director Agent, cinematic shot refinement by the Cinematography Shot Agent, and clip rendering by the Video Generation Agent.Source: Figure 8 · p. 8
  • Human participants rate anonymized, randomly ordered videos against the input script on the same four dimensions, from 1 (very poor) to 5 (excellent).The questions assess script events and characters, camera-operation alignment, visual quality, and resemblance to real-film style.Source: D. Details of Evaluation Details · p. 8

E. Additional Qualitative Comparison.

Additional comparisons contrast Camera Artist’s cinematic shot design and inter-shot continuity with baseline outputs. The framework also supports reference-free storytelling from a textual outline.

E. Additional Qualitative Comparison.

  • Camera Artist’s final wide shot uses a high-angle composition and slow pull-back, while the described baselines lack purposeful or dynamic camera control.The comparison includes Anim-Director’s static framings, VGoT’s medium–long shots without purposeful control, and MovieAgent’s largely static wide shots.Source: E. Additional Qualitative Comparison · p. 8
  • In the Judy truck sequence, Camera Artist preserves character and scene consistency while progressing from discovering the truck to following it; baselines introduce an extra character or an abrupt event jump.Source: E. Additional Qualitative Comparison · p. 8

F. Storytelling without character reference images.

  • Without character reference images, Camera Artist can synthesize characters, scenes, and shot sequences from a textual outline for long-form storytelling.The passage attributes this setting to modern T2I models and multi-reference I2V tools.Source: F. Storytelling without character… · pp. 8–9

G. More Qualitative Results

The supplementary figures and prose present qualitative examples of long-range storytelling, cinematic shot design, and agent prompting. They include a reference-free example generated from a textual story description.

G. More Qualitative Results

  • Scene keyframes and full footage illustrate character identity, spatial continuity, cinematic style, and long-range narrative coherence across complex multi-scene plots.The account also describes smooth shot transitions and consistent visual storytelling.Source: G. More Qualitative Results · p. 9, Figure 14 · p. 13
  • The Director Agent prompt illustration concerns script expansion and scene splitting, while the shot-agent illustration concerns recursive storyboard generation and shot language.Source: Figure 10 · p. 10
  • The VLM-based evaluation prompt figure is described as addressing recursive storyboard generation and the introduction of shot language.Source: Figure 11 · p. 11
  • The comparison figure contrasts Camera Artist’s high-angle, slow-pull-back final wide shot with baseline narrative jumps or irrelevant characters.Source: Figure 12 · p. 12
  • A reference-free figure shows a long-form narrative video generated from a textual story outline without character reference images.Source: Figure 13 · p. 12

Related papers

Loading 2604.09195v1…