acceptodds
Under review as a conference paper at ICLR 2027

FilmFlow: Agentic System for 3D Virtual Film Cinematography

Abstract

Complex camera motion is central to visual storytelling, yet translating cinematographic intent into controllable, executable camera trajectories remains challenging. We introduce FilmFlow, an agentic framework that converts natural-language instructions or reference videos into explicit camera trajectories for dynamic 3D environments. To generate these trajectories, a Director agent constructs executable camera plans by composing high-level motions from concurrent atomic operations that jointly control translation, rotation, and framing. The Director is supported by three specialized agents: an Interpreter that extracts scene-grounded cinematographic intent, a Scout that selects initial viewpoints, and a Critic that evaluates rendered results and provides visual feedback for iterative plan refinement. We also introduce CinemaBench, comprising 150 Blender scenes from self-created and licensed sources, with five camera instructions per scene for 750 text-directed tasks and 60 reference videos associated with a subset of scenes. CinemaBench supports evaluation of both text-directed camera generation and video reference-directed camera generation. Across a diverse set of evaluated baselines, FilmFlow achieves the strongest benchmark performance and the highest user preference, demonstrating effective spatial reasoning, temporal coordination, and visual composition for executable 3D cinematography.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
spatial intelligence

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_2hW7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

FilmFlow presets an agentic system for generating camera trajectories in static and dynamic 3D scenes either from test descriptions or reference videos. The interpreter identifies the main subjects and the camera keywords, the scout searches for an initial viewpoint in conjuction with a critic. The director adds the camera motion and operation timeline. The camera operations are composed of atomic operations where each one can be specified with the corresponding timelines. FilmFlow is evaluated on a new benchmark and compared with several baselines.

Critical strengths
  • The paper tackles a relevant problem where agentic workflows for mapping cinematic intent to camera operations.
  • The separation of an initial view makes sense since this is easier to specify/adjust the camera framing while the camera motion can be added later.
  • A new and good scale benchmark is presented.
Critical weaknesses
  • The proposed method has an initial view planning stage. It might make sense to adapt this to the baselines as well otherwise the transfer of camera trajectories from baselines (LAMP, GenDop, VGGT) is not trivial and it seems often results in the camera looking at empty spaces.
  • Another baseline/ablation can be a single agent instead of the four agents as interpreter/director etc. This can help to strengthen the contribution of the need of four different agents.
  • The method evaluates only on its own Benchmark which might be more tailored for this specific method. Providing results on previously used baselines would be useful.
  • No video results are provided and hence it's not too easy to judge the overall quality of the camera trajectories.
Questions for authors
  • How much of the improvement remains when LAMP and GenDoP receive the same suitable opening as FilmFlow? For reference-video baselines, how are reconstruction scale and placement resolved in the target scene beyond converting coordinate conventions?
  • How does a single agent perform with the same operation library, scene evidence, demonstrations, rendered feedback, and total inference budget?
  • Previous work like Auteur also emphasizes the importance of framing, but also use specific metrics to measure this. Can you report direct camera trajectory metrics, such as collision frequency, framing error etc?
3 comments
AuthorsOpenReview ↗

Response to Reviewer 2hW7

W1/Q1. Shared-opening comparisons and scene-aware V2V placement

We thank the reviewer for highlighting the importance of initialization and for suggesting these controlled comparisons. We provide shared-opening results for LAMP and GenDoP and clarify how the V2V trajectories are placed and scaled in the target scene.

Shared-opening T2V comparison. We rerun LAMP and GenDoP with FilmFlow’s Scout-selected camera position, orientation, and focal length at the same scene-animation frame. Each baseline retains its original motion-generation and translation-scaling procedure and is evaluated on the same test instructions. Higher is better for all metrics.

MethodOpeningAlignment ↑Quality ↑Motion ↑Probe ↑
LAMPScene camera−1.74741.540.2090.564
LAMPShared−1.66742.130.2330.591
GenDoPScene camera−1.76839.630.2570.597
GenDoPShared−1.69244.110.3010.621
FilmFlowScout−1.46147.600.5550.829

Sharing the opening improves both baselines across all four metrics. FilmFlow retains 72.0% of its original Alignment advantage over LAMP (gap: 0.286 → 0.206) and 75.2% over GenDoP (0.307 → 0.231). It also exceeds shared-opening LAMP by 5.47 in Quality, 0.322 in Motion, and 0.238 in Probe, and shared-opening GenDoP by 3.49, 0.254, and 0.208, respectively. These results confirm the contribution of opening-view selection and show that substantial gains remain in the subsequent camera trajectory.

V2V scale and placement. Beyond coordinate conversion, we align each extracted trajectory’s initial pose to the Scout opening and share its focal length and scene-animation start frame. Relative translations are uniformly rescaled using a fixed subject-size-based rule. Relative rotations and temporal structure are preserved, without replanning or per-frame tracking.

The table below compares the original transfer settings with these scene-aware adaptations. Higher is better for all metrics.

MethodPlacementAlignment ↑Quality ↑Motion ↑Probe ↑
LingBot-MapOriginal−1.47235.550.2780.521
LingBot-MapScene-aware−1.30736.110.2900.566
VGGT-ΩOriginal−1.49134.880.1760.533
VGGT-ΩScene-aware−1.35235.920.1650.502
FilmFlowScout−1.22441.770.2890.750

Scene-aware placement improves Alignment from −1.472 to −1.307 for LingBot-Map and from −1.491 to −1.352 for VGGT-Ω. FilmFlow retains the highest Alignment, Quality, and Probe scores, while scene-aware LingBot-Map achieves a slightly higher Motion score (0.290 vs. 0.289). These comparisons show how improved placement benefits trajectory transfer and identify the remaining advantages of FilmFlow’s scene-conditioned planning.

AuthorsOpenReview ↗

Response to Reviewer 2hW7

W2/Q2. Single-agent comparison

We thank the reviewer for suggesting a single-agent comparison to better understand the contribution of FilmFlow’s workflow. We provide results for a single agent using the same backbone and camera-operation representation.

The control uses the same gpt-6-astra model, scene evidence, 22-operation library, executor, and access to the 50-example demonstration bank. One agent selects the opening and generates the complete operation timeline in a single planning pass. This keeps the model and available camera controls consistent while replacing the specialized workflow with direct planning. Higher is better for all metrics.

TaskMethodAlignment ↑Quality ↑Motion ↑Probe ↑
T2VSingle-pass agent−1.53144.260.4310.733
FilmFlow−1.46147.600.5550.829
V2VSingle-pass agent−1.32338.290.2320.701
FilmFlow−1.22441.770.2890.750

FilmFlow outperforms the single-pass agent on all four metrics in both tasks. The absolute gains in Alignment, Quality, Motion, and Probe are 0.070, 3.34, 0.124, and 0.096 for T2V, and 0.099, 3.48, 0.057, and 0.049 for V2V.

These results support the effectiveness of FilmFlow’s complete workflow beyond a single direct planning pass with the same model and operation library. Because this control uses one planning pass, the comparison evaluates the combined contribution of specialized planning and feedback-guided refinement; it does not isolate agent specialization under matched feedback and total inference budgets.

AuthorsOpenReview ↗

Response to Reviewer 2hW7

W3/W4/Q3. External evaluation, direct trajectory metrics, and video results

We thank the reviewer for the helpful suggestions to evaluate beyond CinemaBench, report direct camera-trajectory metrics, and provide full video results. We address these points below.

External evaluation and direct metrics. We additionally evaluate FilmFlow on CinemaTraj’s ScanNet++ scenes and fully specified prompts using its evaluation protocol. This complements CinemaBench with an established external setting.

Motion MSE measures deviation from a hand-crafted reference trajectory after arc-length resampling. CLaTr measures text–trajectory similarity. Collision rate is the fraction of sampled camera positions inside scene geometry; occlusion rate is the fraction of samples at which the target object is blocked. These metrics assess trajectory fidelity, geometric feasibility, and target visibility.

CLaTr is reported as cosine similarity multiplied by 100. Collision and occlusion rates are fractions in [0, 1].

MethodMotion MSE ↓CLaTr ↑Collision rate ↓Occlusion rate ↓
LAMP9.54222.1160.1030.559
GenDoP7.74119.4610.2090.580
CinemaTraj1.74128.9820.0560.503
FilmFlow1.35828.5420.0340.478

FilmFlow achieves the lowest Motion MSE, collision rate, and occlusion rate among the compared methods. Relative to CinemaTraj, it reduces Motion MSE from 1.741 to 1.358 (22.0%), collision rate from 5.6% to 3.4%, and occlusion rate from 50.3% to 47.8%. CinemaTraj retains a slightly higher CLaTr score (28.982 vs. 28.542). These results provide direct evidence of improved trajectory fidelity and fewer collisions and occlusions on an external benchmark.

The geometric metrics complement the video-based evaluation: collision measures camera feasibility, while occlusion measures target visibility. Compositional framing error is a distinct property, so we interpret the occlusion results specifically as evidence about visibility.

Video results. We thank the reviewer for highlighting the importance of viewing the complete camera trajectories. Full rendered videos for both text-directed and reference-video-directed comparisons are included in the supplementary material. Please open demo.html in the supplementary package to browse the comparisons. The videos show how camera motion and subject framing evolve throughout each shot, complementing the sampled frames presented in the paper.

Reviewer_t5PiOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper addresses conditional 3D virtual cinematography that generates camera trajectories in a given 3D scene from an input condition (either a natural language instruction or a reference video).

The main contribution is an agentic framework involving four agents: 1) a director that constructs a camera plan by composing camera motions from atomic operations; 2) an interpreter that extracts scene-based cinematographic intent; 3) a scout the selects initial viewpoints; 4) a critic that evaluates rendered results and offers feedback. A benchmark is also constructed to support the evaluation of text-based and reference-video-based camera motion generation.

Critical strengths
  1. 3D virtual cinematography is an important research problem to study, and the whole system proposed in the paper is well designed, where design choices are justified well.

  2. The proposed system is thoroughly evaluated in the experiments, and achieves strong performance.

Critical weaknesses

The technical novelty is limited. The main contribution of the paper is mainly to combine several agents to a working system. While a lot of engineering effort is involved, the technical novelty of the proposed method is not quite significant.

Questions for authors

What is the size of the example bank in Section 3.3?

1 comment
AuthorsOpenReview ↗

Response to Reviewer t5Pi

W1/Q1. Technical contribution and example-bank size

We thank the reviewer for the positive assessment of the system design and evaluation, and for the opportunity to clarify the technical contribution and example-bank configuration.

Technical contribution. FilmFlow organizes cinematographic planning around two coupled decisions: where a shot should begin and how the camera should move from that opening. Its methodological contribution combines this separation with an executable representation of concurrent camera operations and visual feedback at both stages.

The Interpreter grounds the requested cinematographic intent in the target scene. The Scout uses rendered opening views and the Critic’s feedback to select the initial pose and focal length. With this opening fixed, the Director constructs a timeline of independently parameterized operations whose intervals can overlap. The Critic then evaluates the executed shot and guides revisions to specific operations, parameters, or intervals. This representation allows translation, rotation, and subject-framing controls to operate concurrently.

Figures 9 and 10 illustrate how the two feedback stages address different cinematographic errors. In Fig. 9, the initial portrait already crops the subject’s hair, and the requested fixed-focal-length push-in would tighten the crop further. The Scout selects a wider opening that leaves room for the move. In Fig. 10, the opening correctly frames the platform attendant, but the attendant leaves the frame during lateral camera motion. The Director adds an Aim Follow operation during that interval, preserving the requested camera path while correcting the evolving framing.

The Sequential-only ablation directly evaluates concurrent scheduling: prohibiting overlap reduces T2V Alignment from −1.461 to −1.547. This provides empirical support for the operation representation. The contribution therefore lies in how cinematographic decisions, concurrent controls, and visual feedback are formulated and coordinated within the executable workflow.

Example-bank size. The bank in Section 3.3 contains 50 validated demonstrations, collected from development scenes disjoint from the test scenes. After the Scout selects the opening, FilmFlow retrieves at most one demonstration for the Director. Examples are ranked by the Jaccard similarity between their cinematographic keywords and those of the current task, with ties broken by the Critic score. If no example shares a keyword with the task, the Director proceeds without a demonstration.

Reviewer_2bbAOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents FilmFlow, an LLM-agent system that converts a text instruction or a reference video into an executable camera trajectory in a static or animated Blender scene. Four agents (Interpreter, Scout, Director and Critic) share one proprietary model, gpt-6-astra. The Director writes timelines of 22 parameterised operations whose intervals may overlap, and the Critic guides up to three refinement rounds. The authors also introduce CinemaBench (150 scenes, 750 text instructions, 60 reference clips) and report gains over LAMP, GenDoP, CinemaTraj and pose-transfer baselines on VideoReward, DOVER and a CameraBench probe, together with a 42-participant user study.

Critical strengths

The plan representation is practical and interpretable. Each operation has parameters and an explicit interval, overlap is allowed, and the executor combines relative changes by type, which makes compound moves such as orbiting while rising and pushing in expressible and editable. Critiquing both the opening view and the executed motion, with hard collision checks kept separate from the language scores, is a sensible way to detect framing failures.

The user study is convincing. With 42 participants, hidden method names and randomised order, FilmFlow reaches a pass rate of 84.3% against 63.8% for CinemaTraj on T2V, and Top-1 rates of 61.0% (T2V) and 63.2% (V2V) against a 25% chance level. The same-model zero-shot baseline is ranked first in only 2.3% of V2V judgments, which shows that the agent setup contributes beyond the base model. The split is scene-disjoint and costs are reported.

Critical weaknesses

The benchmark is largely generated by the model that runs the agents. 700 of the 750 instructions were generated by gpt-6-astra and filtered by a "model-based review" (App. B.1), and only 30 of the 500 test instructions are human-written (Table 7). Results on the human-written subset should be reported separately.

The baselines are structurally disadvantaged and the most informative controls are missing. GenDoP and LAMP start from the scene camera, are rescaled heuristically and cannot track moving subjects, CinemaTraj was designed for static scenes, and the V2V pose-transfer baselines are applied "without scene-aware placement". A single-pass gpt-6-astra with the same inventory and operation library, and a scene-aware V2V transfer, would be far more informative comparisons.

The automatic metrics and ablations do not establish the individual contributions. VideoReward alignment is negative for every method on Blender renders, and none of the inexpensive geometric measures available in Blender (subject-in-frame ratio, framing error, occlusion, collision, jerk) is reported. Across all ablations T2V alignment moves only from −1.461 to between −1.474 and −1.547, and V2V Probe is higher without ICL (0.766 vs 0.750). The test split has 46 dynamic and 54 static scenes, but no result separates them, so the claim about dynamic-scene cinematography is not evaluated separately.

The agent framework follows established agent patterns (generate-and-test, critique-and-revise, retrieval of examples), and the closest prior work already composes parametric camera primitives.

Questions for authors
  1. How do FilmFlow and each baseline perform on the 30 human-written test instructions alone, and were the instructions generated with the operation list in the prompt?
  2. What does a single-pass gpt-6-astra achieve with the same scene inventory and operation library, and how does a scene-aware V2V transfer compare?
  3. What are the results for dynamic and static scenes separately?
3 comments
AuthorsOpenReview ↗

Response to Reviewer 2bbA

W1/Q1. Human-written instructions and benchmark construction

We thank the reviewer for this helpful suggestion and for highlighting the importance of evaluating human-written instructions separately. We provide the requested breakdown below, together with further details on the instruction-generation and review procedures.

CinemaBench contains 50 human-written instructions, including 30 in the held-out test split. FilmFlow achieves the highest scores on all four metrics on these 30 instructions. Its Alignment score is −1.344, compared with −1.698 for LAMP, the strongest baseline on this metric, yielding a gain of 0.354. These results support FilmFlow’s effectiveness in following human-authored cinematographic instructions. FilmFlow also leads on all four metrics on the remaining 470 model-generated test instructions.

Instruction sourceMethodAlignment ↑Quality ↑Motion ↑Probe ↑
Human-written (n = 30)LAMP−1.69840.670.2210.549
GenDoP−1.75440.050.2100.601
CinemaTraj−1.80143.890.0090.587
FilmFlow−1.34448.240.5410.787
Model-generated (n = 470)LAMP−1.75041.600.2080.565
GenDoP−1.76939.600.2600.597
CinemaTraj−1.78744.70−0.0730.620
FilmFlow−1.46847.560.5560.832

Instruction-generation prompt. To clarify the information available during generation, the generator received scene evidence, difficulty constraints, motion-category labels, and writing examples. The prompt did not include the complete definitions or parameter interfaces of FilmFlow’s 22 executable operations. We provide a condensed illustration below:

Write camera instructions grounded in the supplied scene and existing animation, with a clear opening, coordinated movement, and deliberate ending. For example:

“Create one continuous ten-second shot at a fixed focal length. Begin high beside the theater sign in a slightly canted close view. Crane down and tilt toward the entrance, revealing the pedestrians and parked coupe. Gradually level the horizon and sweep across the boulevard to pick up the moving green sedan. Track with the car as median trees slide through the foreground, ending with it clearly framed against the opposite shops. Preserve existing animation.”

Use the example’s writing style without copying unsupported subjects or actions.

Review procedure. Regarding the model-based review mentioned in Appendix B.1, programmatic checks first verified object references, difficulty constraints, and timing. A separate gpt-6-astra call then reviewed scene consistency and prose–metadata agreement. Drafts that failed these checks were revised and rechecked.

AuthorsOpenReview ↗

Response to Reviewer 2bbA

W2/Q2. Single-pass control and scene-aware baseline adaptation

We thank the reviewer for these constructive suggestions. To address the concerns about baseline comparability, we provide a single-pass control using the same model and operation library, shared-opening T2V comparisons, and scene-aware V2V transfer results.

Single-pass control. A single gpt-6-astra agent receives the same scene evidence, 22-operation library, executor, and access to the 50-example demonstration bank. It selects the opening and generates the complete operation timeline in one planning pass. FilmFlow outperforms this control on all four metrics in both tasks. The gains in Alignment, Quality, Motion, and Probe are 0.070, 3.34, 0.124, and 0.096 for T2V, and 0.099, 3.48, 0.057, and 0.049 for V2V. These results support the effectiveness of the complete FilmFlow workflow beyond direct planning with the same backbone and executable representation.

Shared-opening T2V comparisons. Following the reviewer’s suggestion, we provide LAMP and GenDoP with FilmFlow’s Scout-selected opening pose and focal length at the same scene-animation frame. Each baseline retains its original motion-generation and translation-scaling procedure. Both improve across all four metrics, confirming the contribution of opening-view selection. FilmFlow nevertheless remains ahead on all four metrics, showing that its advantage extends to the subsequent camera trajectory.

Scene-aware V2V transfer. We adapt the trajectories extracted by LingBot-Map and VGGT-Ω to the target scene by aligning their initial poses to the Scout opening, sharing its focal length and animation start frame, and uniformly rescaling relative translations using a fixed subject-size-based rule. This adaptation preserves relative rotations and temporal structure, without replanning or per-frame tracking.

Results of the additional controls. “Shared opening” uses FilmFlow’s opening pose and focal length at the same scene-animation frame. “Scene-aware” additionally rescales relative translations as described above. Higher is better for all metrics; bold marks the best result within each task.

TaskMethodPlacementAlignment ↑Quality ↑Motion ↑Probe ↑
T2VLAMPScene camera−1.74741.540.2090.564
LAMPShared opening−1.66742.130.2330.591
GenDoPScene camera−1.76839.630.2570.597
GenDoPShared opening−1.69244.110.3010.621
CinemaTrajOwn opening−1.78844.65−0.0680.618
Single-pass agentOwn opening−1.53144.260.4310.733
FilmFlowScout−1.46147.600.5550.829
V2VLingBot-MapOriginal−1.47235.550.2780.521
LingBot-MapScene-aware−1.30736.110.2900.566
VGGT-ΩOriginal−1.49134.880.1760.533
VGGT-ΩScene-aware−1.35235.920.1650.502
Single-pass agentOwn opening−1.32338.290.2320.701
FilmFlowScout−1.22441.770.2890.750

Scene-aware placement improves Alignment from −1.472 to −1.307 for LingBot-Map and from −1.491 to −1.352 for VGGT-Ω. FilmFlow retains the highest Alignment, Quality, and Probe scores, while scene-aware LingBot-Map achieves a slightly higher Motion score (0.290 vs. 0.289). These comparisons show that improved placement reduces part of the original gap, while FilmFlow maintains advantages in instruction alignment, visual quality, and the camera-behavior probe.

AuthorsOpenReview ↗

Response to Reviewer 2bbA

W3/W4/Q3. Scene-type results, direct trajectory metrics, and technical contribution

We thank the reviewer for the helpful suggestions on scene-type analysis, direct trajectory evaluation, and clarification of the methodological contribution.

Dynamic and static scenes. FilmFlow leads on all four metrics in each subset below. On dynamic T2V scenes, its Alignment advantage over the strongest baseline is 0.223. On dynamic V2V scenes, it exceeds the single-pass agent by 0.125 in Alignment, 3.69 in Quality, 0.094 in Motion, and 0.036 in Probe. T2V includes 230 instructions from 46 dynamic scenes and 270 from 54 static scenes; V2V includes 16 dynamic and 24 static reference–scene pairs. Transfer baselines use their original settings.

SubsetMethodAlignment ↑Quality ↑Motion ↑Probe ↑
T2V dynamicLAMP−1.73340.930.1870.548
GenDoP−1.81038.050.2680.571
CinemaTraj−1.91943.18−0.1550.576
FilmFlow−1.51046.060.5470.851
T2V staticLAMP−1.75942.060.2280.578
GenDoP−1.73240.980.2480.619
CinemaTraj−1.67645.900.0060.654
FilmFlow−1.41948.910.5620.810
V2V dynamicLingBot-Map−1.60432.420.2660.577
VGGT-Ω−1.61433.110.2120.508
Single-pass agent−1.37035.910.1800.665
FilmFlow−1.24539.600.2740.701
V2V staticLingBot-Map−1.38437.640.2860.484
VGGT-Ω−1.40936.060.1520.550
Single-pass agent−1.29239.880.2670.725
FilmFlow−1.21043.220.2990.783

Alignment scores. We report VideoReward’s standardized score, (raw TA − μ) / σ. The mean μ and standard deviation σ are fixed by the pretrained checkpoint, not computed from CinemaBench. Consequently, CinemaBench scores need not be centered at zero: all methods can have negative averages when their raw TA averages fall below μ. Zero represents this reference mean, not a threshold for successful instruction following. Higher scores indicate better alignment under the same evaluator.

Direct trajectory evaluation. We use CinemaTraj’s ScanNet++ scenes, fully specified prompts, and evaluation protocol. Motion MSE measures deviation from a hand-crafted reference trajectory after arc-length resampling. CLaTr is text–trajectory cosine similarity multiplied by 100. Collision and occlusion are the fractions of samples with the camera inside geometry and the target blocked, respectively.

MethodMotion MSE ↓CLaTr ↑Collision ↓Occlusion ↓
LAMP9.54222.1160.1030.559
GenDoP7.74119.4610.2090.580
CinemaTraj1.74128.9820.0560.503
FilmFlow1.35828.5420.0340.478

Compared with CinemaTraj, FilmFlow reduces Motion MSE by 22.0%, collision from 5.6% to 3.4%, and occlusion from 50.3% to 47.8%, while CinemaTraj retains a slightly higher CLaTr score. Together, these results extend the evaluation beyond CinemaBench and show improvements in reference-trajectory fidelity, collision avoidance, and target visibility.

Component contributions. We further evaluate the three component ablations under the same external protocol.

VariantMotion MSE ↓CLaTr ↑Collision ↓Occlusion ↓
w/o Scout1.49425.6880.0370.526
w/o Critic1.56224.2610.0390.550
w/o ICL1.43926.8290.0360.507
FilmFlow1.35828.5420.0340.478

Removing any component degrades all four metrics. Removing the Critic produces the largest degradation, increasing Motion MSE from 1.358 to 1.562 and occlusion from 47.8% to 55.0%. These results support opening-view selection, visual critique, and example retrieval.

On CinemaBench, removing ICL raises V2V Probe from 0.750 to 0.766, as the reviewer notes. Retaining ICL improves Alignment from −1.296 to −1.224, Quality from 39.04 to 41.77, and Motion from 0.172 to 0.289.

Technical contribution. We appreciate the reviewer’s observation that critique-and-revise, retrieval, and parametric primitives are established ideas. FilmFlow integrates them with concurrent operation scheduling, opening-view selection, and feedback on executed motion. Whereas CinemaTraj sequentially composes trajectory segments for static scenes, FilmFlow assigns operations independent, overlapping intervals for simultaneous translation, rotation, and framing control with moving-subject adaptation. Prohibiting overlap in the Sequential-only ablation reduces T2V Alignment from −1.461 to −1.547, supporting the concurrent representation. The single-pass comparison further evaluates the complete workflow using the same backbone, scene evidence, and executable operation library.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.