acceptodds
Withdrawn from ICLR 2027

Navigational Compositional Guidance: Learning Content-Adaptive Video Trajectories for Photography

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Traditional image composition methods rely on basic geometric transformations and classical aesthetic rules for optimization. With the development of multimodal models, numerous effective composition recommendation methods have been proposed recently. However, most existing methods merely provide static image-based composition references with simple text instructions, which cannot deliver refined and practical guidance for real shooting scenarios. Video-based composition guidance is more consistent with real-world dynamic rules due to video's inherent spatial-temporal continuity. Nevertheless, current video composition methods suffer from limited generalization for complex composition tasks, only supporting basic camera and perspective adjustments, and cannot meet fine-grained guidance demands. To tackle these drawbacks, this paper proposes a novel video model-based content-adaptive composition path recommendation algorithm, named Navigational Composition Guidance. It achieves complex-scene dynamic composition guidance and realizes autonomous navigation within the high-dimensional aesthetic manifold to search for optimal composition paths.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Computational Photography, Composition Optimization, Camera Motion Generation, Video Generation, Aesthetic Assessment, Reinforcement Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_GtPSOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Navigational Compositional Guidance (NCG), a video generation model that provides dynamic camera trajectory guidance to help users achieve professional compositions. Key contributions include a synthetic data generation pipeline (using outpainting and virtual cameras), a two-stage training paradigm (SFT and GRPO), and a direction-aware metric (DAPE). The method is validated through automated metrics and human evaluations.

Critical strengths
  1. Introducing GRPO to video composition is effective. It successfully addresses the "one-to-many" mapping issue and the lack of a stopping mechanism, empowering the model to autonomously seek aesthetic optima and achieve terminal settling.

  2. The authors solve the extreme scarcity of expert kinetic data by "reverse-engineering" trajectories. Using image outpainting and virtual camera projection on high-quality static photos generates realistic camera motion videos without expensive manual data collection.

Critical weaknesses
  1. The evaluation relies heavily on metrics directly optimized during training. The model is explicitly trained to maximize the PQA reward via GRPO, yet PQA is used as a primary comparative metric in Table 2, violating independent evaluation principles. Similarly, the DAPE metric is trained using temporally reversed videos as negative samples and then evaluated on its ability to detect reversed videos. Even if the scenes do not overlap, testing on the exact artificial task seen during training evaluates a learned shortcut rather than true aesthetic direction awareness.

  2. As a generative model, NCG must "hallucinate" out-of-frame content when starting from a truncated view (e.g., half a face). In real-world photographic guidance, the physical scene will likely not match the AI's hallucinated pixels. This fundamental disconnect between generative hallucination and physical reality severely undermines the practical utility of the proposed "guidance."

Questions for authors
  1. Since DAPE was explicitly trained to penalize temporally reversed videos, testing its accuracy on reversed videos seems circular. How does DAPE perform on independent, non-reversed video pairs of varying compositional quality?

  2. Regarding the out-of-frame hallucination issue, how do you propose to resolve the practical misalignment when the user moves the physical camera, but the real-world scene does not match the AI's hallucinated pixels?

2 comments
AuthorsOpenReview ↗

To GtPS. This response addresses Critical Weakness 1 and Question 1 under “Questions for Authors.”

To GtPS. This response addresses Critical Weakness 1 and Question 1 under “Questions for Authors.”

We thank the reviewer for highlighting the distinction between reward-aligned evaluation and independent evidence, as well as the limitations of reversal-based testing.

1. Evaluation beyond the optimized PQA reward.

We agree that PQA scores are not independent of our GRPO objective and should be interpreted as reward-aligned measurements. However, our evaluation also includes VBench 2.0, DeQA, Q-Align, and human assessments, which are not used as the GRPO reward.

These evaluations provide complementary evidence: VBench 2.0 assesses image quality, aesthetic quality, motion smoothness, and dynamic degree, while DeQA and Q-Align provide additional image-based assessments. These dimensions are related to, but not interchangeable with, PQA’s VQ, MQ, and CA scores.

2. DAPE training includes negatives beyond temporal reversal.

Temporal reversals are one type of negative sample, rather than the sole source of supervision for DAPE. Training also includes inferior-endpoint samples, as described in Appendix C, Table 9. Our training design follows the PQA framework and adds reversal-based supervision to encourage sensitivity to the direction of compositional change.

The motivation is documented in the newly added Appendix D.4: inspections of individual video pairs and the reported aggregate results show that PQA CA scores change little when frame order is reversed. This motivates explicit directional supervision.

AuthorsOpenReview ↗

To GtPS. This response addresses Critical Weakness 2 and Question 2 under “Questions for Authors.”

To GtPS. This response addresses Critical Weakness 2 and Question 2 under “Questions for Authors.”

We acknowledge that generating a camera trajectory from a single image may introduce synthesized out-of-frame content, even when that content appears realistic and plausible. Such content is not guaranteed to match the physical scene, and this is a limitation of generative approach. Some recent state-of-the-art composition optimization methods, such as PPC and PhotoFramer, also introduce out-of-frame hallucinations to some extent.

Our objective is to provide a visual reference for composition adjustment, emphasizing aspects such as subject placement, relative scale, and spatial balance. For short-range adjustments, we expect some discrepancies in incidental scene details to be tolerable when the suggested compositional change remains applicable to the real scene.

Our formulation assumes a static scene, which reduces complications arising from independently moving objects but does not eliminate uncertainty about unseen content. Accordingly, we intend the generated video to illustrate a possible compositional adjustment, rather than serve as a faithful prediction of the scene that the user will observe after moving the camera.

Therefore, although hallucinations may have some impact, we believe that they neither create a fundamental disconnect from physical reality nor substantially undermine the practical utility of the proposed guidance.

Reviewer_s5JvOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the generation of video trajectories that progressively improve photographic composition from a single input image. The proposed Navigational Compositional Guidance (NCG) framework constructs eight categories of camera-motion data using 2D virtual-camera transformations and generative viewpoint editing/video interpolation, then adapts HunyuanVideo-1.5 in two stages. Supervised fine-tuning gradually replaces motion-specific prompts with a generic composition-improvement prompt. Reinforcement learning subsequently uses SAGE-GRPO with PQA rewards for visual quality, motion quality, and compositional aesthetics. The paper also introduces DAPE, a learned evaluator of the direction of compositional change.

Validation includes automatic image/video metrics, training-stage ablations, human preference rankings, DAPE direction discrimination on 400 independent scenes, and an adaptation experiment using 100 pairs of favorable and perturbed compositions. The results support compositional improvement and adaptation of motion intensity on the evaluated scenes. However, the independent contribution of the proposed training design and the validity of parts of the trajectory evaluation require clarification.

Critical strengths
  1. Content adaptation is supported by meaningful independent human evidence. Appendix H.3.2 and Table 18 compare favorable and perturbed inputs from 100 independent scene pairs, with human initial–final comparisons that conceal method identities and input conditions. Full NCG achieves 73% non-degradation on favorable inputs and 82% improvement on perturbed inputs, compared with 55% and 51% without GRPO. This experiment is more informative than reward increases or reduced terminal optical flow alone: it supports an improvement in task-relevant behavior rather than indiscriminate motion suppression.
  2. The integration of compositional motions and direction-sensitive evaluation is useful. The data pipeline covers four elementary and four compound motions, providing a concrete approach to constructing training resources for video-based composition guidance. DAPE's forward–reverse evaluation on 400 independent scenes controls frame content, uses human-verified labels and candidates from multiple generation methods, and avoids selection by DAPE scores. Its 98.25% direction accuracy provides useful evidence of direction sensitivity, complementing conventional image aesthetics and video quality metrics.
Critical weaknesses
  1. The key new design is not isolated by the necessary controls, leaving the contribution unclear. PPC already studies video-based composition guidance and preference optimization. The PQA evaluator used here is also adopted from PPC and serves in data filtering, GRPO rewards, and evaluation, while SAGE-GRPO supplies the video post-training framework. PQA itself is therefore not a new contribution, although using these existing components is not inherently a weakness. What needs to be established is the additional benefit of the proposed data covering elementary and compound motions and the progressive prompting strategy, including whether these designs improve adaptation to initial composition quality. However, the “without SFT/without GRPO” ablations in Tables 4 and 15 only establish the usefulness of the two stages; they do not establish that progressive prompting outperforms training with a generic prompt throughout. Without this direct comparison under the same backbone, data, and training budget, and without an untuned-backbone baseline, it is difficult to distinguish benefits of the proposed training design from those of task-specific data combined with existing post-training methods. This directly affects my assessment of methodological novelty and contribution.
  2. The evaluation does not fully substantiate the scope of the claims about composition optimization and camera-motion realism. DAPE's independent reversal test is a strength, but without an initial–final aesthetic-score-difference baseline, it does not establish the added value of evaluating the complete path over evaluating endpoints. Human results mitigate this concern, but Table 14 uses only 18 input scenes without scene-level uncertainty estimates; Table 18 also lacks uncertainty estimates for paired differences in human-assessed outcomes between methods. More concretely, the OOD classifier in Appendix G takes only camera positions and rotations while including a zoom-in category, although ideal optical zoom changes intrinsics rather than extrinsics. Its endpoint-displacement normalization may also amplify noise for near-zero translation. Consequently, the 94.46% “Legal Rate” is a detector acceptance rate, not direct evidence of physical executability. These issues limit the scope of the supported performance claims; they do not demonstrate wholesale failure of the method.
Questions for authors
  1. Evidence for the key design: With backbone, data, and training budget fixed, how much does progressive prompting improve over using the generic prompt throughout? Are corresponding results available after SFT and after GRPO to distinguish the effect of prompting from task-specific data and post-training? This comparison would directly affect my assessment of the methodological contribution.
  2. Evaluation validity: How does DAPE compare with an initial–final aesthetic-score-difference baseline? Can the existing human ratings support scene-level paired uncertainty estimates? For the trajectory detector, please explain how extrinsics alone handle pure optical zoom and how normalization and detection remain reliable with near-zero endpoint translation. If the 94.46% result denotes only distributional acceptance, please narrow the associated claims accordingly.
2 comments
AuthorsOpenReview ↗

To s5Jv. This response addresses Critical Weakness 1 and Question 1 under “Questions for Authors.”

To s5Jv. This response addresses Critical Weakness 1 and Question 1 under “Questions for Authors.”

We thank the reviewer for identifying this gap in contribution attribution. We agree that the original “without SFT/without GRPO” ablations establish the usefulness of the training stages but do not isolate progressive prompting. We have therefore expanded Table 4 and updated the corresponding discussion in the main text to include an untuned HunyuanVideo baseline, unified-prompt SFT, and unified-prompt SFT followed by GRPO.

1. Controlled comparison of progressive and unified prompting

The unified-prompt baseline uses the generic instruction throughout SFT. The unified and progressive variants use the same backbone, SFT data, and training budget, followed by the same GRPO configuration and budget. This comparison isolates the prompting strategy while holding task-specific training data and the post-training procedure fixed.

After SFT, progressive prompting increases the Legal Rate from 76.03% to 92.03%, a gain of 16.00 percentage points, compared with unified prompting. VQ/MQ/CA also increase from 0.3665/0.3508/0.3714 to 0.3946/0.3905/0.4349. These results show that the progressive curriculum provides additional benefits beyond training on the task-specific data with a generic prompt.

After the same GRPO stage, these benefits persist. Progressive prompting increases the Legal Rate from 89.45% to 94.46%, a gain of 5.01 percentage points, while VQ/MQ/CA increase from 0.4511/0.4751/0.4790 to 0.4947/0.4941/0.4828. The strongest improvements are in trajectory acceptance and visual quality; the additional CA improvement after GRPO is comparatively modest. Legal Rate and OOD rejection rate are complementary statistics and should not be interpreted as independent evidence.

Together, the comparisons before and after GRPO demonstrate a measurable contribution from progressive prompting beyond the use of SFT itself or the existing GRPO framework.

2. Untuned-backbone baseline and contribution attribution

The newly included untuned HunyuanVideo baseline achieves 51.84% Legal Rate and 0.3659/0.3511/0.3328 VQ/MQ/CA. Unified-prompt SFT increases Legal Rate to 76.03% and CA to 0.3714, while VQ and MQ remain approximately unchanged. Progressive-prompt SFT further improves all of these metrics. Including these conditions helps distinguish the effect of task-specific fine-tuning from the additional benefit of the prompting curriculum.

We also agree that PQA is not a new contribution of this work. We adopt PQA from PPC and build on the video post-training framework provided by SAGE-GRPO. Our claimed contributions concern the construction of composition-guidance data covering elementary and compound motions and the progressive training strategy for learning guidance from a generic instruction, rather than the invention of these existing evaluators or optimization frameworks. The new prompting controls establish the curriculum's benefit on the same training dataset.

3. Adaptation to initial composition quality

Table 18 provides complementary evidence about the final model's adaptation to initial composition quality: NCG achieves 73% non-degradation on favorable inputs and 82% improvement on perturbed inputs, compared with 55% and 51%, respectively, without GRPO.

However, this comparison isolates the effect of GRPO within our training pipeline; it does not establish that progressive prompting specifically causes better adaptation to initial composition quality. We therefore use Table 18 to support the final model's content-dependent behavior, and the expanded Table 4 to establish the additional benefit of progressive prompting under controlled training conditions.

We appreciate the reviewer's suggestion. These additions make the contribution attribution more precise and provide the direct comparisons missing from the original submission.

AuthorsOpenReview ↗

To s5Jv. This response addresses Critical Weakness 2 and Question 2 under “Questions for Authors.”

To s5Jv. This response addresses Critical Weakness 2 and Question 2 under “Questions for Authors.”

We thank the reviewer for identifying these evaluation limitations. We have added the requested baseline and clarify the scope of our evidence below.

1. Endpoint baseline for DAPE.

We added a DeQA-based initial–final aesthetic-score-difference baseline to Appendix D.3, including its definition and results. On the same 400 human-verified video pairs, it achieves 95.50% direction accuracy, compared with 98.25% for DAPE, a gain of 2.75 percentage points. This supports DAPE’s improved direction discrimination over the tested endpoint baseline.

2. Uncertainty in human evaluations.

We limited Table 14 to 18 scenes to manage participant workload and fatigue. We acknowledge the limited scene coverage and have not completed scene-level paired uncertainty analyses for either Table 14 or Table 18. We therefore report these results as descriptive evidence on the evaluated scenes. We will explicitly acknowledge this limitation and incorporate paired uncertainty estimates in future evaluations.

3. Zoom-in representation and normalization.

We apologize for the ambiguity in the term “zoom-in.” In our design, this category denotes a dolly-in approximation, where the camera moves forward to obtain a closer composition, rather than pure optical zoom through a focal-length change. We do not introduce a dedicated optical-zoom parameter; the intended movement is represented by forward translation in the estimated camera positions. Thus, extrinsic are used to characterize this translational motion, not to recover optical zoom. Dolly-in and optical zoom are not physically equivalent, and we will clarify this distinction and the category terminology in Appendix G.

Regarding normalization, the detector uses endpoint displacement with an additive ϵ\epsilon, which prevents division by zero but does not establish robustness when displacement is small. We acknowledge that pose-estimation noise may be amplified in this regime. The accumulated-path-length normalization used separately for MS does not resolve this detector limitation, and we will state this explicitly.

4. Meaning of “Legal Rate.”

We thank the reviewer for prompting us to clarify the purpose and interpretation of this metric. Directly verifying the physical camera motion underlying a generated video is challenging. We therefore designed the trajectory detector as a proxy for assessing whether the estimated camera trajectory is consistent with the reference distribution of the predefined motion categories.

Accordingly, the reported 94.46% denotes the acceptance rate under the specified detector and preprocessing pipeline. We interpret this as indirect evidence of consistency with the reference motion patterns, rather than a direct verification of the underlying physical camera motion.

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

NCG generates videos illustrating composition adjustments from an input photograph using a generic composition-improvement prompt. It constructs synthetic trajectories covering four atomic camera motions and four combinations through virtual-camera transformations and generative viewpoint interpolation. HunyuanVideo-1.5 is adapted with progressive prompt replacement and PQA-based GRPO. The paper also trains a direction-aware evaluator, DAPE. Evaluation includes automatic metrics, human quality judgments and a held-out favorable/perturbed-input study. These experiments assess generated content and motion adaptation; they do not test users following the guidance to take photographs.

Critical strengths
  1. Direct evidence of content-dependent adjustment in generated videos. Table 18 tests 100 held-out scene pairs, includes all generated outputs and uses blinded initial/final-frame judgments. NCG achieves 73% non-degradation on favorable inputs and 82% improvement on perturbed inputs, versus 55%/51% without GRPO. Together with the paired motion measurements, this supports learned adaptation beyond indiscriminate motion suppression, although it does not establish effectiveness for users operating a camera.

  2. A substantial data pipeline with complementary quality checks. The paper combines annotated compositions, four atomic motions and four compound templates, with filtering and human assessment of 150 synthetic videos (B.3). DAPE uses sources disjoint from NCG and achieves 98.25% direction accuracy on 400 independent, human-verified forward/reverse pairs (D.3). These checks provide evidence beyond PQA, which also serves as the filtering and training signal. The DAPE result specifically supports direction discrimination on the evaluated pairs.

Critical weaknesses
  1. Methodological increment and video necessity remain insufficiently established. The approach combines synthetic trajectories, SFT and GRPO; the aesthetic manifold is conceptual (H.3.1). Its four atomic motions and four combinations lack a simple transform-prediction/interpolation control for the 2D subset. Removing SFT also does not isolate progressive prompting from unified-prompt SFT. These gaps leave unclear what the proposed designs add beyond task-specific adaptation and simpler guidance generation.

  2. Guidance utility and computational cost are not directly evaluated. Photography experts participate in B.3/F.2, but judge generated quality; no study measures actual camera use, novice understanding or learning. Table 18 supports generated compositional improvement, without establishing added utility over a target image with text/arrows. Despite using an 8.3B video model (E.1), the paper reports no measured inference latency, peak memory or operating cost, leaving the intended usage setting and benefit–cost tradeoff unclear.

Questions for authors
  1. Contribution attribution: How does equal-budget unified-prompt SFT followed by the same GRPO compare with progressive prompting? For the 2D subset, is a control available that predicts transformation parameters and renders a smooth trajectory? These comparisons would clarify the curriculum's contribution and the need for full video generation.

  2. Usage and evidence: What is the intended usage setting, and what are the measured inference latency and peak memory on specified hardware at the evaluated resolution/frame count? Is there evidence that users follow video guidance more successfully than a target image with text/arrows? A matched user-operation study would be additional evidence; the existing quality-preference studies address a different outcome.

2 comments
AuthorsOpenReview ↗

To rCuU. This response addresses Critical Weakness 1 and Question 1 under “Questions for Authors.”

To rCuU. This response addresses Critical Weakness 1 and Question 1 under “Questions for Authors.”

We thank the reviewer for these constructive suggestions. We agree that removing SFT alone does not isolate the contribution of progressive prompting, and that comparison with simpler trajectory-generation alternatives helps clarify the scope of our approach.

1. Isolating the contribution of progressive prompting.

We have expanded Table 4 and updated the corresponding discussion in the main text to include the original HunyuanVideo baseline, unified-prompt SFT, and unified-prompt SFT followed by GRPO. The unified and progressive variants use the same SFT data and training budget, followed by the same GRPO configuration and budget.

Before GRPO, progressive prompting improves the Legal Rate from 76.03% to 92.03% and the compositional-aesthetics score from 0.3714 to 0.4349, compared with unified-prompt SFT. After GRPO, the progressive variant retains an advantage: the Legal Rate increases from 89.45% to 94.46%, while VQ/MQ/CA increase from 0.4511/0.4751/0.4790 to 0.4947/0.4941/0.4828. These comparisons support a contribution from the prompting curriculum beyond the inclusion of SFT itself, with the clearest gains in trajectory acceptance and visual quality. Legal Rate and OOD rejection rate are complementary statistics, rather than independent evidence.

We also clarify that the aesthetic manifold is a conceptual interpretation of the objective, not an explicitly recovered geometric structure or a theoretical convergence guarantee.

2. Scope of parameter prediction and interpolation for 2D guidance.

Our objective is a unified model for 2D and 3D compositional guidance, requiring only an input photograph and a generic instruction, without user-specified motion types or transformation parameters.

We agree that predicting transformation parameters and rendering an interpolated path is a meaningful baseline for the 2D subset. In particular, zoom-in and shifts or rotations whose sampling regions remain within the input image can be rendered directly. We do not claim that full video generation is necessary for all such cases.

The limitation arises when a trajectory reveals regions outside the input image. For these cases, transformation parameters alone cannot supply the missing visual content; an additional completion mechanism, such as image outpainting or generative view synthesis, is required. True viewpoint changes additionally involve depth-dependent parallax and disocclusion, which cannot generally be represented by a single 2D image transformation. These considerations motivate our use of a video model as a common representation for the broader task, rather than establish its superiority over parameter-based rendering on every 2D example.

3. Additional interpolation analysis and its interpretation.

To investigate the endpoint-interpolation alternative, we have added an experiment described in the appendix, with qualitative video comparisons in the supplementary material. This experiment examines the behavior of endpoint-conditioned interpolation under changes in viewpoint and motion extent.

We distinguish this analysis from the reviewer's proposed parameter-prediction baseline: endpoint interpolation evaluates the temporal synthesis component, but does not by itself isolate the quality of learned transformation prediction or establish the necessity of video generation for in-bounds 2D transformations. We therefore use this experiment to characterize the interpolation alternative and its limitations.

To examine this limitation, we conducted an endpoint-conditioned geometric interpolation experiment, detailed in Appendix I, with qualitative results provided in the Additional_interpolation_study folder of the supplementary materials. The results show that even examples within our 2D subset can exhibit changes in occlusion and relative object positions that are not adequately captured by a global image-plane transformation. Although these adjustments are labeled as zoom-in, shift, or rotation, their realization through camera motion can involve viewpoint changes and, when camera translation is involved, depth-dependent parallax. Consequently, interpolating geometric parameters and blending the endpoint images can produce visible misalignment and ghosting, as illustrated in the supplementary videos. These observations highlight a limitation of the simple transform-and-interpolate alternative: a smooth parameter trajectory does not necessarily yield a visually coherent transition when scene geometry and visibility change.

We appreciate the reviewer's suggestions, which have helped us separate the evidence for progressive prompting from the broader motivation for unified video-based guidance.

AuthorsOpenReview ↗

To rCuU. This response addresses Critical Weakness 2 and Question 2 under “Questions for Authors.”

To rCuU. This response addresses Critical Weakness 2 and Question 2 under “Questions for Authors.”

We thank the reviewer for highlighting the need to clarify both the practical usage setting and the evidence for guidance utility. We have added inference-cost measurements in Appendix J and a novice-user preference survey in Appendix K.

1. Inference cost and intended usage setting.

On an NVIDIA H20 GPU, generating one video takes 327.42 seconds, with a peak GPU memory usage of 14,540.8 MiB (approximately 14.20 GiB) under our evaluation configuration. These measurements are now reported in Appendix J.

We acknowledge that this latency does not support real-time feedback during camera operation. The intended applications of the current implementation are therefore pre-shoot planning, offline reference, and photography learning assistance, where users can review a generated composition-adjustment video before attempting the corresponding adjustments. These are intended use cases; their practical effectiveness remains to be evaluated directly. The computational cost is a current limitation, and the benefit–cost tradeoff for interactive use has not yet been established.

2. Additional evidence of novice-user preference.

Following the reviewer's suggestion, we conducted an anonymous survey with 42 participants without professional photography experience. We prepared eight sets of examples, each showing three guidance formats for the same scene: a scene image with textual instructions, a scene image with directional arrows, and video guidance. For each set, participants selected the single format they preferred overall, jointly considering ease of understanding and perceived ease of following.

Video guidance received 80.65% of valid preference selections, compared with 7.14% for Image + Text and 12.20% for Image + Arrows. The study design and results are provided in Appendix K. This provides additional evidence that novice users prefer video guidance in the evaluated examples.

We agree, however, that preference does not establish more successful camera operation. This survey neither measures actual execution performance nor separately quantifies comprehensibility and ease of following. We explicitly acknowledge these limitations in Appendix K.

3. Matched user-operation study.

We agree that a matched user-operation study would provide more direct evidence of practical utility. Such a study has not been conducted in the current work, and we therefore do not claim that video guidance has demonstrated superior execution accuracy, task completion time, or final photographic quality over static guidance.

In future work, we plan to compare video guidance against a target image with text/arrows under matched scenes and target compositions, measuring these practical outcomes. The current evidence supports the quality and content adaptivity of generated guidance, together with novice users' subjective preference for its video format. Our quantitative evaluations also demonstrate the advantages of NCG in improving the composition of generated outputs over the evaluated baselines. A matched user-operation study would further assess whether these benefits translate into better photographs taken by users following the guidance.

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

Related papers

Discussion (0)

Sign in to comment.