Better Call CineCrew: Consistent Ultra-Long Narrative-to-Film Generation¶
Conference: ECCV2026
Paper: Official paper page · PDF
Area: Video Generation
Keywords: Long-form narrative video generation, FilmDSL, multi-agent systems, cross-clip consistency, persona consistency
TL;DR¶
CineCrew uses FilmDSL to organize narrative intent, cinematic staging, assets, and repair feedback into an executable production workflow without retraining video models, improving character, prop, and story continuity on 20 MovieBench narratives without leading every visual-quality or physics metric.
Background & Motivation¶
Splitting a screenplay into short clips and stitching them together bypasses the duration limit of a single generation call, but it does not solve state inheritance. A character may change identity in the next shot, a prop may disappear, or a room may reset its layout. Even with a stable face, an unusually exaggerated action from a normally cautious character can disrupt the story. MovieAgent and AniMaker already employ multi-agent planning, so the question here is not whether to divide the work, but what representation should constrain the interdependent decisions made after that division.
A screenplay usually describes events without exhaustively specifying shot size, camera placement, prop locations, or an actor's micro-actions. Turning story paragraphs directly into video prompts lets each generation independently fill those gaps. A character reference image helps constrain appearance but cannot fully specify who should look at whom, how an emotion develops, or which objects must persist. Stronger video generators can improve individual clips without supplying an inspectable, locally editable production state across shots.
The paper places its contribution between scripts and generators: narrative intent, cinematographic decisions, and rendering instructions are stored separately, while all production roles share assets and continuity records. Core idea: use FilmDSL as a persistent production protocol, express cross-clip requirements as explicit fields, and repair generation failures locally through keyframe anchoring and structured diagnostic feedback.
Method¶
Overall Architecture¶
The input can be a film plot, an episode summary, or script-like text; the output is an assembled long-form video and its FilmDSL specification. The system first identifies narrative beats, such as an information reveal, an action completion, or an emotional change. A beat can require multiple executable clips, and clip durations can vary, so a beat should not be equated with a fixed-duration shot.
Pre-production establishes global specifications and asset anchors, then compiles clip plans through hierarchical FilmDSL and persona constraints. Production uses keyframe-first execution to invoke existing generators clip by clip. Diagnostic repair and the shared Rulebook feed failures and accepted state back into the workflow, before post-production assembles clips, subtitles, and sound. These constraints are explicit tool conditions and inspection targets, not mathematical guarantees that a generator must satisfy.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
Input["Narrative text"] --> Assets["Global Specifications<br/>and Asset Anchors"]
Assets --> DSL["Hierarchical FilmDSL<br/>and Persona Constraints"]
DSL --> Execute["Keyframe-First Execution"]
Execute --> Review["Diagnostic Repair<br/>and Shared Rulebook"]
Review -->|Revise clip fields and retry| Execute
Review -->|Carry state into subsequent clips| DSL
Review -->|Finalize clips| Output["Post-production<br/>Video, subtitles, and sound"]
Key Designs¶
1. Global Specifications and Asset Anchors: make every clip reference the same characters and sets
The Showrunner specifies frame rate, aspect ratio, tone, era, location, and cast. These global headers are inherited unless a clip explicitly overrides them. The Art Department then constructs Character Sheets and Set Assets: the former provide multi-view identity anchors, wardrobe references, and compact persona information; the latter provide key images or descriptions of primary locations. A clip therefore references an established character and set instead of independently reimagining a generic middle-aged man. Global formatting and assets are part of FilmDSL rather than detached production attachments.
FilmDSL does not embed raw images. It stores stable references such as IDs, URIs, hashes, or embedding keys, allowing downstream roles to access the same assets through fields such as char_id, sheet_ref, and set_ref. This reduces drift introduced by repeated textual descriptions. Asset referencing should be distinguished from semantic retrieval: the paper explains persistent references and access to assets, but does not specify a separately trained retrieval module or ranking algorithm. The assets also support more than visual resemblance, because persona information in character sheets participates in performance compilation.
2. Hierarchical FilmDSL and Persona Constraints: separate events, filming decisions, and generator calls
The unified JSON-like object contains meta, assets, memory, and clips. Each clip preserves three layers instead of collapsing all requirements into a single prompt. Equation (3) expresses this structure as:
Narrative Action, \(a_i\), records observable events, emotion, and dialogue, authored by the Story Editor. Cinematic Staging, \(s_i\), specifies shot size, camera motion, framing, lighting, characters, and props through the Cinematographer. Render Specification, \(r_i\), is compiled by the Technical Director into keyframe prompts, video prompts, optional negative prompts, and tool arguments such as duration, frame rate, aspect ratio, and seed. Keeping these layers makes it possible to distinguish a problem in narrative intent from one in cinematographic choice or rendering instructions, instead of rewriting the entire story whenever a clip fails.
Staging also carries continuity fields, including character_refs, set_ref, required_props, forbidden_props, and continuity_link. The first two bind identity and location; the others identify objects that must exist, objects that must not appear, and relationships to previous clip state. When adjacent clips should behave like a continuous take, last-frame-to-first-frame chaining can be used. This is not a universal requirement at every editing boundary. Distinguishing continuous takes from ordinary cuts helps preserve legitimate film editing instead of imposing visual continuity everywhere.
Appearance anchors alone cannot ensure that a character behaves in character. The Persona Schema therefore combines a persistent trait-to-behavior mapping with a beat-level performance block. The former translates traits such as caution or impatience into posture, gaze, reaction latency, interpersonal distance, and gesture rhythm. The latter specifies the current performance objective, local emotion arc, blocking intent, and micro-actions. The Acting Coach instantiates the block from the current beat and continuity state, and the Technical Director incorporates it into staging constraints and renderable prompts. Abstract personality can thus influence observable behavior, without implying that the paper trains a dedicated acting model.
3. Keyframe-First Execution: establish identity, layout, and props before adding motion
The Production Operator first generates a storyboard-like keyframe using the clip specification, character references, set assets, and locked layout. This converts textual constraints into a joint spatial arrangement: where the character stands, whether the required prop is present, and whether the lighting matches the intended look. The video generator then receives the keyframe and clip prompt. Compared with a character-only reference, the keyframe also anchors the current composition and staging; compared with inheriting only the previous final frame, it can accommodate a new camera position or scene arrangement.
Execution has dialogue and non-dialogue branches. For dialogue, the VO Director synthesizes a voice track from FilmDSL dialogue and supplies it to an audio-video-conditioned generator. Without dialogue, the video generator consumes the keyframe and video prompt. The cached extraction of Equation (2) is damaged, but Section 3.4 and the algorithm explicitly describe both branches, so this note preserves the verified mechanism in words rather than guessing the broken equation. The VO role also handles vocal intent and timing. The Post Supervisor assembles clips, subtitles, Foley, and sound effects, with optional ambience or music mixing; the audio pipeline is not separately ablated quantitatively.
4. Diagnostic Repair and Shared Rulebook: localize failures to fields and retain their corrections
The Dailies Reviewer checks identity drift, prop duplication or teleportation, layout resets, and camera-motion mismatch, producing structured QA signals. Repairs prioritize continuity and constraint fields, such as required or forbidden props, continuity links, and locked entities. They then modify keyframe prompts, followed by video prompts, and revisit upstream staging if necessary. This is more actionable than assigning an entire video a low score: an unexplained object appearance first calls for a targeted constraint revision, not replanning the complete narrative through every crew role.
The Production Rulebook combines static priors with dynamic state. Static information includes staging heuristics, anti-hallucination and anti-spawning constraints, and naming conventions. Dynamic entries include accepted keyframes, locked props and layouts, failure tags, and finalized prompts. Relevant constraints are selectively injected into the Story Editor, Cinematographer, Technical Director, Production Operator, and Dailies Reviewer, while execution logs update their shared state. The algorithm permits at most \(R\) attempts per clip and stops early on acceptance. The main text does not specify the value of \(R\), acceptance thresholds, or a dedicated fallback after exhausting retries. This memory is production state rather than a parameter update, and its ability to grow indefinitely without errors is not established.
A Worked Example¶
The paper's Gene narrative includes kneading dough in a bakery, returning to a dark apartment, and playing a VHS recording. The following connects those events using the documented mechanism; it is an explanatory walkthrough, not an additional experimental trace.
Global specifications establish the black-and-white noir style and character appearance, while the asset library stores Gene's references and relevant sets. The Story Editor determines the current action and emotion rather than merely stating that he is anxious. The Persona Schema makes that anxiety observable through gaze and movement. The Cinematographer chooses how to show the character and surroundings, and the Technical Director compiles these decisions into executable clip specifications.
On returning to the apartment, the system references the same character asset and the apartment set, using a keyframe to establish the character, recording equipment, and room layout. If the generated video makes the equipment disappear, the Reviewer flags prop permanence, revises the relevant fields, and requests another attempt without changing the story objective of playing the recording. Accepted visual state and constraints enter the Rulebook for subsequent television and reaction shots. The paper does not report a specific failure count or retry gain for this walkthrough, so no success rate is invented.
Loss & Training¶
This is an orchestration system that does not retrain its base generators; it introduces no new diffusion loss, optimizer, or learned continuity objective. Improvements come from structured planning, asset conditioning, generator calls, and diagnostic retries. Consequently, deployment cost should account for model calls and regeneration rather than only training expenditure. The cached main text does not fully enumerate the generator stack, per-clip inference configuration, total cost per narrative, or retry distribution; those hyperparameters should not be filled in speculatively.
Key Experimental Results¶
Main Results¶
Automatic evaluation samples 20 MovieBench narratives across genres, periods, indoor and outdoor settings, day and night, cast sizes, and action- or dialogue-driven stories. CineCrewEval separates cinematic expression, cross-clip consistency, and general VBench quality. Its VLM judge is Gemini-3-Flash, and the LTX-Studio baseline uses LTX-2 Pro. The following selects columns from the paper's Table 2; scores with different units should not be added together.
| Method | B.R. Beat Readability | P.A. Persona Alignment | N.C. Narrative Coherence | I.C. Identity Consistency | P.P. Prop Permanence | C.L. Cinematic Look | P.C. Physics Compliance |
|---|---|---|---|---|---|---|---|
| MovieAgent | 2.50 | 3.20 | 3.00 | 60.09 | 58.78 | 4.10 | 3.35 |
| AniMaker | 3.20 | 3.60 | 4.10 | 61.38 | 65.17 | 4.20 | 3.50 |
| LTX-Studio | 4.25 | 4.10 | 3.80 | 67.15 | 73.24 | 4.90 | 4.60 |
| CineCrew | 4.90 | 4.75 | 4.70 | 75.12 | 78.59 | 4.85 | 4.20 |
B.R. assesses whether key actions, intentions, and emotions are readable; P.A. assesses behavior against traits and motivations; N.C. assesses temporal and causal continuity. I.C. concerns character identity across clips, and P.P. concerns props appearing, disappearing, or duplicating without justification. C.L. covers lighting, color, texture, and style, while P.C. checks teleportation, spawning, and impossible motion. The main text provides these semantic definitions but refers detailed evaluation rules to supplementary material absent from the current cache. I.C. and P.P. therefore should not be reinterpreted as precisely defined detection accuracies.
Ablation Study¶
The following selects columns from the paper's Table 4; higher is better throughout. Removing FilmDSL means directly decomposing scripts into prompt-ready clips, not merely removing the JSON serialization wrapper.
| Configuration | B.R. | P.A. | N.C. | I.C. | P.P. | L.C. Layout Continuity |
|---|---|---|---|---|---|---|
| Without Showrunner | 4.65 | 4.05 | 3.95 | 67.82 | 70.15 | 4.20 |
| Without Production Rulebook | 4.55 | 4.70 | 3.90 | 73.78 | 76.85 | 4.25 |
| Without Assets Memory | 4.05 | 3.50 | 4.05 | 61.29 | 72.72 | 4.05 |
| Without Dailies Reviewer | 4.75 | 4.75 | 4.60 | 74.08 | 70.24 | 3.95 |
| Without FilmDSL | 3.05 | 3.40 | 3.45 | 65.24 | 69.15 | 4.00 |
| Full system | 4.90 | 4.75 | 4.70 | 75.12 | 78.59 | 4.50 |
The user study recruits 24 participants and uses 10 narratives, each spanning at least 5 consecutive clips. Outputs are anonymized and randomly ordered, and the paper states that generation budgets are matched. The next table selects criteria from the paper's Table 3, rated from 1–5.
| Method | CS Cinematic Staging | BC Beat Clarity | CC Cross-clip Continuity | OP Overall Preference |
|---|---|---|---|---|
| MovieAgent | 2.67 | 3.35 | 2.17 | 3.67 |
| AniMaker | 3.33 | 2.83 | 4.33 | 3.00 |
| LTX-Studio | 4.00 | 3.67 | 3.83 | 3.17 |
| CineCrew | 3.67 | 4.50 | 4.83 | 4.25 |
Key Findings¶
- Against LTX-Studio, I.C. improves by 7.97 points and P.P. by 5.35 points, placing the main advantage in cross-clip state preservation. These are score differences, not relative percentages or claims of statistical significance.
- Removing FilmDSL reduces B.R. by 1.85 points and N.C. by 1.25 points, the largest drops for both metrics in the ablation table, supporting the importance of explicit narrative intent. Removing the Production Rulebook also reduces N.C. to 3.90, showing that shared state contributes to narrative coherence.
- Removing Assets Memory reduces I.C. by 13.83 points, the largest identity loss among the ablations. Removing the Reviewer reduces P.P. by 8.35 points, suggesting distinct roles for visual anchoring and failure repair.
- The method does not win everywhere: LTX-Studio scores 4.90 on C.L. and 4.60 on P.C., above CineCrew's 4.85 and 4.20. It also leads user-study CS with 4.00 against 3.67. These mean-score trends should be read alongside the absence of confidence intervals.
Highlights & Insights¶
- FilmDSL matters because failures can be traced to narrative, staging, or rendering decisions, not because JSON itself is novel. A similar layered interface could support illustrated stories or advertising videos, letting a human director change only the failed aspect of the expression.
- Character consistency is separated into stable appearance and behavior appropriate to the role. Mapping enduring traits into observable actions, then combining them with the current beat's performance requirements, is more inspectable than repeatedly appending personality adjectives to prompts.
- The asset library records what entities should be, while the Rulebook records what has happened and which attempts failed. Together they support cross-clip production state, although the ablations demonstrate utility only in the evaluated setting, not unlimited memory capacity.
Limitations & Future Work¶
- The cached main paper has no dedicated limitations section; these are assessment points based on its reporting scope. Despite the ultra-long framing, it does not clearly summarize total generated duration, maximum clip count, or performance as sequence length increases. Quantitative validation of hour-long filmmaking should not be inferred.
- Automatic evaluation uses 20 narratives, while the user study involves 24 people and 10 narratives, without confidence intervals or significance tests. Cinematic expression relies heavily on a single VLM judge. Detailed judging prompts and identity or prop metric formulas are absent from the current cache, limiting independent verification.
- Different systems may use different generators. Although the user study states that budgets are matched, the main text does not fully specify common call costs or execution configurations. Main-table gains cannot be attributed entirely to orchestration; controlled ablations are closer to that question but still lack variance over repeated trials.
- The Persona Schema, speech branch, and repair priorities are not independently ablated. Retry limits, acceptance thresholds, and memory-growth costs are also insufficiently reported. Future evaluation should jointly measure gains, call expenditure, latency, and the downstream propagation of repairs.
- Explicit constraints do not eliminate the generator's physical errors. Stronger spatial-state checks and human intervention interfaces are promising directions, but these are proposed extensions rather than capabilities already implemented and validated here.
Related Work & Insights¶
- Compared with MovieAgent: both borrow production-team organization. CineCrew emphasizes a persistent action–staging–render representation and field-level repair. Its evaluated system is more consistent, but MovieAgent should not be reduced to a single unplanned prompting call.
- Compared with AniMaker: AniMaker emphasizes multi-agent animated storytelling and MCTS-driven clip generation. CineCrew focuses on carrying characters, props, layouts, and performance intent across clips through FilmDSL, rather than only searching for a better candidate clip.
- Compared with FilMaster and LTX-Studio: FilMaster informs the evaluation design through cinematic language and post-production timelines; LTX-Studio is the commercial system actually compared in the main table. CineCrew is stronger on narrative and continuity, while LTX-Studio retains advantages on some visual and physical dimensions. These two systems should not be conflated.
Rating¶
- Novelty: 4/5. An integrated executable film representation, behavioral constraints, and persistent repair state, built on existing multi-agent and generator toolchains.
- Experimental Thoroughness: 3/5. Multiple baselines, component ablations, and a user study are valuable, but long-horizon scale, cost, variance, and evaluation reproducibility remain underreported.
- Writing Quality: 4/5. The three-layer representation and production loop are clear; some implementation details are deferred to supplementary material, and broad claims require checking against the tables.
- Value: 4/5. Useful guidance for controllable multi-shot production, without establishing that system-level improvements solve filmmaking at arbitrary lengths.