FilmAgent: A Multi-Agent Framework for End-to-End Film Automation in Virtual 3D Spaces

Framework

• FILMAGENT automates end-to-end virtual film production by assigning director, screenwriter, actor, and cinematographer roles across three sequential workflow stages.The stages are idea development, scriptwriting, and cinematography, with role-specific collaboration throughout.Source: §1 · pp. 2–3

Evaluation

• 3.98 out of 5 average human-evaluation score across 15 ideas shows FILMAGENT outperforming single-agent efforts on plot coherence, dialogue-profile alignment, camera appropriateness, and actor-action accuracy.The evaluation covers four aspects and reports consistent gains for the collaborative system.Source: §1 · p. 2
• FILMAGENT uses coordinated multi-agent collaboration to exceed the performance of the more advanced single-agent o1 model despite using GPT-4o as its foundation.The comparison is presented as evidence for the value of coordination beyond the choice of foundation model.Source: §1 · p. 2

Discussion

• Compared with Sora, FILMAGENT offers coherent, physics-compliant storytelling in predefined 3D spaces, whereas Sora adapts more readily to diverse scenes but struggles with consistency and narrative delivery.The systems therefore exhibit complementary strengths and weaknesses in filmmaking.Source: §1 · p. 2

Collaboration

• Its two collaboration strategies iteratively refine scripts and camera settings, substantially reducing hallucinations and improving production quality.Critique-Correct-Verify is used in scriptwriting, while Debate-Judge is applied in cinematography.Source: §1 · pp. 2–3

2.1 VIRTUAL FILM PRODUCTION

• Virtual film production covers computer-aided production and visualization methods, including narrative design, camera control, cutting, and editing.Its virtual nature also supports remote collaboration and accessibility.Source: §2.1 · p. 3
• Deep-learning approaches have automated virtual cinematography through styled camera-motion generation and LSTM-based keyframing for finer control.These methods operate in computer-graphics environments and use video references or keyframes.Source: §2.1 · p. 3
• LLM-based virtual-production systems have addressed text-driven motion scheduling, extracting character movement sequences from textual descriptions.Story-to-Motion is an example of using LLMs to map long text into text, position, and duration tuples.Source: §2.1 · p. 3

2.2 MULTI-AGENT FRAMEWORK

• Single agents often struggle with effective, coherent, and accurate problem solving when meaningful collaborative interaction is required.This limitation motivates the transition toward multi-agent frameworks.Source: §2.2 · p. 3
• Multi-agent research improves reasoning and factuality through agent discussions and simulates standard human practices and workflows.This line of work includes simulated societies and collaborative systems such as ChatDev and MetaGPT.Source: §2.2 · pp. 3–4
• FILMAGENT applies multi-agent collaboration to filmmaking within a Unity-built virtual space with pre-configured actor positions and camera setups.The living-room example includes static shots at multiple distances and dynamic shots that follow or orbit characters.Source: Figure 2 · p. 4

3.1 ENVIRONMENT SETUP

• Configured 3D spaces provide 15 locations, predefined actor positions, 21 actions, nine shot types, and synchronized speech, actions, and camera timing.The environment includes 32 standing points, 33 sitting points, and 165 static plus 107 dynamic shots.Source: §3.1 · pp. 4–5

3.2 OVERVIEW

• FILMAGENT defines four specialized roles—Director, Screenwriter, Actor, and Cinematographer—to divide filmmaking into focused tasks.The director oversees profiles, outlines, feedback, discussions, and final decisions; the screenwriter specifies dialogue, positioning, and actions.Source: §3.2 · p. 5

3.3 AGENT COLLABORATION STRATEGIES

• Critique-Correct-Verify generates, critiques, corrects, and verifies responses iteratively until the critique agent approves the result.The process can stop early when the updated response is judged sufficiently high quality.Source: §3.3 · p. 6
• Debate-Judge has peer agents independently propose and critique responses before a judgment agent delivers the final decision.In cinematography, two cinematographers debate discrepant camera choices and the director participates in the final adjudication workflow.Source: §3.3 · pp. 6–7

3.4 WORKFLOW

• Scriptwriting explicitly annotates each dialogue line with character positioning and actions, then refines the script through director and actor feedback.The director-screenwriter cycle addresses plot coherence and actions, while actors check consistency with character profiles before finalization.Source: Figure 4 · p. 7, §3.4 · pp. 7–8
• The framework operates in three sequential stages: idea development, scriptwriting, and cinematography.The director expands a brief idea into character profiles and scene outlines, collaborators refine dialogue and movements, and cinematographers assign camera choices.Source: §3.4 · p. 7

4.2 RESULTS

• 3.98 out of 5 average score: FILMAGENT configurations consistently outperform the standalone CoT agent across human-evaluated film qualities.Table 2 evaluates actor actions, plot coherence, script alignment with actor profiles, and camera-setting appropriateness.Source: §4.2 · p. 8, Table 2 · p. 9
• FILMAGENT’s iterative feedback and revisions produce significant improvements across all aspects, especially plot coherence and camera-setting appropriateness.The framework decomposes complex tasks into manageable sub-tasks and compares Solo and Group configurations to assess collaboration benefits.Source: §4.2 · pp. 8–9
• Despite using GPT-4o, multi-agent FILMAGENT outperforms the single-agent o1, showing that coordination can exceed the performance of a more advanced underlying model.The o1-based CoT agent outperforms GPT-4o-based CoT and surpasses single-agent FILMAGENT in some aspects, but not the multi-agent framework.Source: §4.2 · p. 9

4.3 PREFERENCE ANALYSIS

• Human evaluators prefer revised scripts and camera choices after collaboration, with examples showing fewer hallucinated actions, better character alignment, and more suitable, varied shots.The analysis covers 15 scripts and 50 camera-choice modifications using Critique-Correct-Verify and Debate-Judge strategies.Source: §4.3 · p. 9

Comparison with Sora

• Sora’s short-duration generation is evaluated by assigning 3–4 seconds per scene within a storyboard capped at 10 seconds.FILMAGENT’s planned scenes provide the prompts for each Sora video segment.Source: §5 · pp. 10–11
• Sora adapts quickly to diverse scenes, realistic styles, and shots, whereas FILMAGENT emphasizes coherent, physics-compliant videos with stronger storytelling.The comparison presents the systems as complementary rather than interchangeable.Source: §5 · p. 11, Figure 6 · p. 11

Limitations and Future Work

• FILMAGENT relies on predefined virtual 3D spaces with limited action spaces and preset camera settings, constraining fine-grained control and flexibility.Future directions include adaptable scene synthesis, motion and camera adjustment, multimodal LLMs, and expanded crew roles.Source: §5 · p. 11

6 CONCLUSION

• FILMAGENT is an LLM-based multi-agent framework for end-to-end film production in virtual 3D spaces.It combines constructed 3D spaces, simulated human workflows, and multi-agent collaboration strategies.Source: §6 · p. 12
• 3.98 out of 5 average human-evaluation score supports FILMAGENT’s effectiveness for generated videos.Further analysis reports improved script quality, camera selection, and reduced hallucination errors.Source: §6 · p. 12

A ENVIRONMENT DETAILS

• 15 locations comprise the constructed virtual 3D spaces used for film production.The locations are documented through screenshots in Figure 7.Source: A ENVIRONMENT DETAILS · p. 15
• The environment includes nine annotated camera-shot types with descriptions and views documented in the paper’s tables.Both static and dynamic shots are represented.Source: A ENVIRONMENT DETAILS · p. 15
• Characters can perform 21 predefined actions from the Mixamo website, including talking, arguing, walking, and emotional gestures.The action library defines the character behaviors available in the constructed environment.Source: A ENVIRONMENT DETAILS · p. 15

• Theme fidelity and actor-action accuracy are evaluated with binary yes/no judgments.Source: B HUMAN EVALUATION GUIDELINES · p. 15
• Annotators rate camera-setting appropriateness on a 1–5 Likert scale.Source: B HUMAN EVALUATION GUIDELINES · p. 15
• Script alignment with actor profiles and overall plot coherence are also rated from 1 to 5.Source: B HUMAN EVALUATION GUIDELINES · p. 15

• The annotated script excerpt is ready for simulation and filming in the constructed 3D virtual spaces.Source: C SCRIPT EXAMPLE · p. 15
• Walking actions document both initial and destination actor positions because walking changes position.Source: C SCRIPT EXAMPLE · p. 15
• Walking is represented separately in the JSON file’s "move" dictionary.Source: C SCRIPT EXAMPLE · p. 15

Workflow overview

• FILMAGENT’s workflow comprises three stages, with prompts mapped to each stage and their usage in Table 6.Source: D PROMPT LIST · p. 15, Table 6

Idea development

• Idea development prompts generate character profiles and coherent scene plans from a provided topic.Source: D PROMPT LIST · pp. 21–22, Figure 10 · p. 21, Figure 11 · p. 22

Scriptwriting

• Scriptwriting prompts add dialogue, character positions, actions, and movement annotations to each scene.Source: D PROMPT LIST · pp. 23–24, Figure 12 · p. 23, Figure 13 · p. 23

Action annotation

• Action selection uses predefined actions, state constraints, one action per character at a time, and dialogue-based emotional reasoning.Source: D PROMPT LIST · pp. 24–25

Script refinement

• Director-screenwriter refinement follows a Critique-Correct-Verify cycle covering plot coherence, action appropriateness, revision, and verification.Source: D PROMPT LIST · pp. 21, 26, 27, Figure 16 · p. 26, Figure 17 · p. 27

Feedback and cinematography

• Actors provide profile-based feedback, while cinematographers debate camera choices and the director resolves conflicts to finalize setups.Source: D PROMPT LIST · pp. 21, 28, Figure 19 · p. 28