Photography Perspective Composition: Towards Aesthetic Perspective Recommendation

We sincerely thank the reviewer for the careful review and constructive comments. Below, we have provided a point-by-point response. The "Q" stands for question and the "W" stands for weakness. Numbers in parentheses (e.g., W1(1)) represent sub-questions of that question.

Q1 & W1(1). About the dataset detail.

We are willing to provide detailed information about our dataset. Our final dataset contains 100K high-quality perspective transformation videos, split into 90K/5K/5K for training/validation/testing. Specifically, our automated pipeline builds upon the approach described in Section 3.2. First, we start with expert images from multiple professional photography datasets, including GAIC, SACD, FLMS, FCDB, and Unsplash, totaling approximately 90K expert-composed images. Second, we adopt ViewCrafter to generate video sequences that transition from optimal to suboptimal perspectives, then reverse these sequences to obtain our training data. Third, PQA is applied to filter the data.

We will release the dataset and complete project code to promote this research area.

Q2 & W1(2). Red bounding boxes and homography transformations.

We would like to clarify that these components, including the red bounding boxes, feature matching, and homography transformations are not part of the learning process but rather serve as auxiliary guidance tools for practical user interaction.

Specifically, our implementation works as follows: We first draw a guidance box on the final optimal perspective frame, which is a rectangular box centered at the frame center. Then, using feature matching between the initial and final perspectives, we extract corresponding point pairs and estimate the homography matrix. We then apply this homography transformation to project this box onto the original image, creating a distorted box shape. As users physically move their camera following our video guidance, this box gradually changes shape and approaches a rectangular form when reaching the optimal perspective.

W1(4). "Ours" model identification.

Unlike typical papers that focus on comparing specific models, we want to clarify that in our experiments, "ours" refers to our PPC pipeline rather than any specific model.

In our tables, we highlight the Wan2.1 14B model in gray to indicate it as our primary experimental configuration. We select this model based on its superior performance in our evaluations, and it is used to generate visualization results.

Our paper presents a pipeline rather than modifications to the video generation model, where video generation models serve as implementation tools. We extensively tested three SOTA video generation models (CogVideoX 1.5 5B, Hunyuan I2V, and Wan2.1 14B) to validate our pipeline.

Q4 & W2. About the human preferences and the inter-annotator reliability in PQA.

We appreciate the reviewer's concern. We provide detailed information about our human preference collection and inter-annotator reliability.

Our PQA model training involves 8 expert annotators with professional photography backgrounds. For standard videos, each video is annotated by 3 annotators using pairwise comparison across three dimensions: Visual Quality (VQ), Motion Quality (MQ), and Composition Aesthetic (CA). The video pairs in our annotations consist of videos generated by different video models using the same input image.

To assess inter-annotator reliability, we employ Fleiss' Kappa, a statistical measure designed to evaluate agreement among multiple raters:

We can observe that our annotations demonstrate substantial consistency across three metrics.

About the Expert Bias.

To address this, we conduct additional validation with broader user groups. Due to time constraints, we conduct a preliminary validation study with 28 amateur users who evaluated 20 randomly selected video pairs across three dimensions through questionnaire (different from the one in the rebuttal for Reviewer 4nUR). Expert annotations for these videos are determined by majority vote from our 3-expert panel.

The agreement analysis shows that amateur users aligned well with expert consensus: 91.2% agreed with experts on VQ, 85.8% on MQ, and 82.3% on CA (averaged across 20 videos). These agreement rates demonstrate reasonable alignment between expert and amateur preferences.

We plan to expand this validation study to include a larger and more diverse group of amateur users in future work.

Q5 & W3. About the visual attribute changes.

We thank the reviewer for this important observation. We would like to clarify our core motivation and our contribution to the visual attribute change.

Real-World Perspective Changes. In real-world photography scenarios, perspective transformations naturally introduce visual attribute changes, such as varying lighting conditions, altered subject-background relationships, and changes in depth of field. These changes are inherent to the physical nature of perspective adjustment. While these visual attribute variations are important to consider, our core motivation is to establish a perspective-based compositional pipeline that guides users in improving their shots through optimal perspective adjustments.

Even so, during the model training process, we have made several efforts to minimize visual attribute distortions:

(1) Angle Limitation Strategy. During data generation, we limit the maximum angle difference (20°) between the first and last frames of training videos. In our initial experiments without this limitation, the model often exhibited severe hallucinations by generating completely new scenes with drastically different lighting and exposure conditions.

(2) RLHF-Based Consistency Optimization. Upon analyzing these cases, we discovered that while some videos exhibited attribute changes, their perspective transformations actually aligned well with human aesthetic preferences. This led us to introduce a second-stage RLHF training that relaxes rigid constraints and allows the model to learn from human preference feedback.

Quantitative Analysis of Visual Attribute Changes. To assess the visual attribute variations in our generated videos, we conduct measurements across key visual dimensions. Our analysis evaluated 1,200 videos:

These results demonstrate that our PPC methodology maintains reasonable visual attribute stability, and our additional control mechanisms provide further refinement.

Furthermore, we are actively developing the Unreal Engine 5-Based Data Generation Pipeline. This game engine approach provides significantly more controlled environments for generating training data. Through perfect lighting control, we can eliminate the lighting inconsistencies that commonly occur in real-world 3D reconstruction by providing precise environmental lighting control.

W1(3). RLHF implementation clarification.

The reviewer's question about "prompts used" reflects a misunderstanding of our RLHF approach. Our RLHF implementation does not rely on text prompts, as the operational logic and purpose differ from typical text-based RLHF applications. Instead, our RLHF operates directly on image-to-video generation models, optimizing the perspective transformation quality.

For defining higher-quality and lower-quality videos, we clearly specify in our paper that win/lose video pairs are determined by our PQA model, which serves this dual purpose of both data filtering and RLHF preference determination. The PQA model evaluates videos across three dimensions (visual quality, motion quality, and composition aesthetic) to establish preference rankings, eliminating the need for manual preference annotations in the RLHF process.

W4(1). About the trajectory.

The "trajectory" mentioned in Line 120 refers to the camera motion trajectories used in our automated dataset construction (Section 3.2 in the paper).

W4(2). About the shared linear projection head.

We appreciate this insightful question about our architectural choice. While the reviewer's intuition about separate projection heads for different reward types is reasonable, our shared linear projection head design follows the causal attention mechanism. The true separation occurs at the attention level, not the projection level. Each token has already acquired fundamentally different feature representations through causal attention constraints. Therefore, a single shared linear head is sufficient to map these distinct token representations to their respective dimension scores.

We sincerely thank the reviewer for the positive feedback and constructive suggestions on our paper. Here are our responses to the reviewer's comments. We use different letters to represent different types of comments, where "W" represents weakness and "Q" represents question.

Q1. About the camera motion trajectory generation.

Our camera trajectories are not purely random but incorporate systematic design principles. We have explored and experimented with multiple trajectory generation strategies across different complexity levels.

(1) Basic Single-Direction Movements. We first established fundamental directional movements using ViewCrafter's parameter system. Our basic movements include four primary directions: left and right movements involve horizontal angle adjustments while maintaining vertical position, creating lateral camera motion for reframing subjects. Up and down movements adjust the vertical angle while keeping horizontal position constant, allowing for elevation changes that can improve horizon placement or alter subject prominence within the frame.

(2) Compound Dual-Direction Movements. Building on basic movements, we explore diagonal combinations that provide more sophisticated compositional adjustments. These compound movements simultaneously adjust both horizontal and vertical angles to create diagonal camera trajectories. Upper-left and upper-right movements combine upward motion with lateral adjustments, particularly useful for horizon placement optimization and diagonal composition rebalancing. Lower-left and lower-right movements blend downward motion with horizontal shifts, helping to improve subject-background relationships and achieve perspective correction in architectural or landscape photography.

(3) Complex Multi-Phase Movements. We also experiment with complex motion patterns for advanced compositional scenarios. Our experiments systematically explore different complexity levels of camera movements. (1) We employ 3-step patterns (40% of cases) that sequentially combine horizontal, vertical, and diagonal movements to address fundamental compositional imbalances through systematic perspective correction. (2) We utilize 4-step patterns (30% of cases) that enhance the basic sequences by adding zoom adjustments to the 3-step movements, enabling simultaneous optimization of both framing and perspective. (3) We also implement 5-step patterns (30% of cases) that were built upon the 4-step sequences by randomly adding one additional movement type (horizontal, vertical, diagonal, or zoom), providing enhanced flexibility for complex compositional refinements.

We evaluate different trajectory strategies using metrics from our experiments:

However, for amateur users, overly complex camera motion trajectories may be difficult to understand and reproduce. Therefore, we adopt a simplified hybrid approach combining three types of movements. We primarily utilize single-step movements (40% of cases), two-step movements (40% of cases), and three-step movement sequences (20% of cases). This distribution ensures that the guidance remains straightforward while still accommodating more sophisticated compositional needs when required.

W1. Comparison to cropping-based methods.

We conduct comparisons from two aspects: user study and VILA aesthetic evaluation.

We conduct a user study with 57 amateur users through questionnaire. For other detailed designs, please refer to our response to Reviewer 4nUR's Q2(1).

(1) We prepared 12 test cases covering diverse photography contexts (portrait, landscape, architecture, multi-subject scenes). For each scenario, we applied two SOTA cropping methods (Gencrop (AAAI24) and Cropper (CVPR25)) and selected the highest-confidence crop result as the baseline comparison.

C1. Overall composition improvement compared to the best cropping result:
    Rate PPC performance: 1 (Much worse), 2 (Worse), 3 (Similar), 4 (Better), 5 (Much better)

C2. Content preservation compared to cropping methods:
    Rate PPC performance: 1 (Much worse), 2 (Worse), 3 (Similar), 4 (Better), 5 (Much better)

Here are the results:

The results suggest that users generally prefer our PPC method over cropping approaches. In particular, users appreciated that PPC preserves more image content (88% high satisfaction, mean 4.6/5) compared to cropping, which removes parts of the image. The overall improvement scores (78% high satisfaction, mean 4.2/5) indicate that perspective-based recomposition could be a promising alternative to traditional cropping.

(2) To provide objective validation beyond subjective user preferences, we also employ VILA to evaluate 800 image pairs comparing PPC perspective transformations (last frame) against best-performing cropping results from Gencrop and Cropper. VILA evaluated both PPC-transformed images and crop-enhanced images on the same 0-1 aesthetic quality scale. The percentage improvements in aesthetic scores compared to original images are categorized into different ranges:

These results demonstrate that our perspective-based recomposition method provides superior performance compared to traditional cropping-based approaches.

W2. About using synthesis-based approach.

Thanks for the reviewer's concern. Synthesis-based approaches do suffer from inherent limitations, including synthesis artifacts and input-output content inconsistency.

Despite these limitations, our core motivation addresses scenarios where cropping-based methods cannot provide adequate solutions. In many compositional challenges, cropping is inherently insufficient because it operates solely within the 2D image plane and cannot address spatial relationship issues or perspective distortions.

To mitigate these consistency issues, we have implemented several technical strategies:

(1) We restrict perspective transformations to controlled angles (20°) to maintain visual consistency and reduce structural alterations. This constraint significantly minimizes the likelihood of severe hallucinations while preserving meaningful perspective guidance.

(2) Our RLHF approach aligns generated outputs with human aesthetic preferences, reducing artifacts through preference-based optimization rather than forcing strict ground truth adherence.

(3) We are developing Unreal Engine 5-based data generation pipelines for more controlled, higher-quality training data that will fundamentally reduce synthesis artifacts.

W3. Comparison with the direct image-to-image approach.

We sincerely appreciate this thoughtful suggestion. We initially explored image-to-image (I2I) approaches, but encountered two fundamental challenges that led us to abandon this direction:

(1) Guidance Limitation. I2I approaches often generate images with excessive angular changes, producing completely different perspectives that users find difficult to interpret for practical guidance. When users are presented with only the initial and final images, they struggle to understand the specific camera movements or positioning adjustments needed to achieve the transformation in real-world scenarios.

(2) Training Challenges. We initially experimented with I2I approaches but encountered problems when applied to perspective transformations. During training, the models frequently converged to simply reproducing the original input image, failing to learn meaningful perspective adjustments. This convergence issue makes it difficult to achieve reliable perspective transformation outputs. Additionally, even when the models do learn to transform perspectives, they often generate inconsistent results with severe artifacts and content distortions.

Our core idea focuses on providing instructional guidance rather than just a final result. Users need to understand how to achieve the transformation in real scenarios, which requires demonstrating the step-by-step perspective adjustment process. The video-based approach naturally provides this educational component by showing the gradual transformation sequence, enabling users to internalize the movement patterns and apply them in their own photography practice. As video generation models continue advancing, this approach becomes increasingly viable and effective.

We sincerely thank the reviewer's feedback and constructive suggestions. Here are our responses to the reviewer's comments. The "W" represents weakness, "Q" represents question, and numbers in parentheses (e.g., Q2(1)) represent sub-questions of that question.

Q1 & W2. About the data ratio influence.

We are continuing to conduct experiments and have found that the core issue may lie in the significant gap between our perspective transformation data and the original pretraining data of video generation models. Most of the original pretraining data consists of movie clips and internet videos, which differ substantially from our perspective transformation videos in terms of motion patterns and visual characteristics.

Since the original pretraining data of video generation models is not publicly available, we use VBench's (CVPR24) test videos as a proxy to approximate the original distribution. VBench provides a comprehensive video quality benchmark with diverse video samples. We compare our perspective transformation data against this VBench distribution to quantify the distributional gap:

As dataset diversity increases, all distribution divergence metrics worsen significantly. This distributional shift forces the model to deviate from its learned priors.

Q2(1) & W1 & W4. User usability assessment.

We conduct a preliminary user survey to validate our method's practical value. Due to time constraints, our survey scale is not very large but focus on gathering meaningful feedback about the system's usability and effectiveness.

(1) Questionnaire Design.

We design and distribute an online questionnaire, and we receive 57 users' feedback. The design of the questionnaire is as follows:

A. Background Assessment

A1. Photography Experience:
    □ Beginner (<6 months)         □ Novice (6-24 months)
    □ Intermediate (2-5 years)     □ Advanced (>5 years)

A2. Primary Equipment:
    □ Smartphone only (iPhone, Samsung, Vivo, Google Pixel)
    □ Point-and-shoot camera (Ricoh GR III, Fujifilm X100V)
    □ DSLR/Mirrorless (Canon R5, Sony A7IV, Fujifilm X-T5)
    □ Professional gear (Hasselblad, Leica)

B. PPC System Evaluation

We set up 8 diverse scenario cases to evaluate the system's performance across different photography contexts. For each scenario, users evaluate the guidance videos generated by PPC.

*(7-point Scale: 1=Strongly Disagree, 4=Neutral, 7=Strongly Agree)*

B1. The final perspective significantly improves upon the original photo
B2. The transformation sequence is intuitive and easy to follow
B3. The suggested compositions align with my personal aesthetic preferences

C. Comparative Assessment with Crop-Based Methods

We prepared 12 test cases covering diverse photography scenarios. For each case, we applied two SOTA cropping methods (Gencrop (AAAI24) and Cropper (CVPR25)) and selected the highest-confidence crop result.

C1. Overall composition improvement compared to the best cropping result:
    Rate PPC performance: 1 (Much worse), 2 (Worse), 3 (Similar), 4 (Better), 5 (Much better)

C2. Content preservation compared to cropping methods:
    Rate PPC performance: 1 (Much worse), 2 (Worse), 3 (Similar), 4 (Better), 5 (Much better)

(2) Results.

A. Participant Profile Analysis: Our participants had diverse photography experience: 34% beginners, 44% novice, 19% intermediate, and 3% advanced. For equipment, 41% used smartphones, 25% point-and-shoot cameras, 31% DSLR/mirrorless, and 3% professional gear.

B. System Evaluation Results (7-point Scale)

C. Comparative Assessment Results

Q2(2). First vs last frame.

In our user study described above, we conduct first vs. last frame comparison. Specifically, item B1 "The final perspective significantly improves upon the original photo" received positive feedback: mean score of 5.6/7.0, with 75% of participants rating high satisfaction (6-7 points) regarding the aesthetic improvement from initial to final frames.

To further validate these findings on a larger scale, we also conduct aesthetic assessment using VILA on our test set. Since this evaluation covers multiple diverse videos rather than individual videos, we employed aesthetic improvement ratio as our primary metric.

The results demonstrate consistent aesthetic improvements across our PPC. We plan to conduct more extensive experiments with a larger and more diverse user base for further validation.

Q3. About the low-quality inputs.

We thank the reviewer for their comments and have conducted additional experiments accordingly. We used Q-Align (ICML24), a commonly used Image Quality Assessment model, to identify and categorize low-quality images.

Q-Align provides quality scores on a 0-5 scale. Our experiment images consist of two main categories: high-quality and low-quality images. The high-quality subset contains 300 images with Q-Align scores of 3 or higher. The low-quality subset comprises 300 images with Q-Align scores of 2.0 or lower.

These results show that while there is some performance degradation on low-quality inputs, PPC still maintains acceptable performance levels compared to all inputs.

Q4. About the implementation.

To provide users with more intuitive and clear guidance, we adopt the video generation approach to implement our PPC, which is a novel direction that has not been explored before in photography composition guidance. However, typical video generation models can take 5 minutes to generate a few seconds of video, mainly due to model size, diffusion steps, and frame count requirements.

To address this inference time issue, we make several optimization attempts. Most importantly, we observed that our perspective transformation task does not require as many frames, diffusion steps, or high resolution as conventional video generation. By reducing the frame count to 9 frames, using only 10 diffusion steps (typically 50), and generating at a lower resolution, we achieved over 60x speedup while maintaining effectiveness for our specific use case. For the CogVideoX 1.5 5B model, the inference time is 5.1 seconds with 10 GB GPU memory usage under BF16 precision.

With further optimization for deployment engineering and the rapid advancement of video generation models, PPC will become faster and more suitable for everyday photography use.

Q5. Analysis with existing cropping-based methods.

We conduct comparisons from two aspects: a user study and VILA aesthetic evaluation. Our comprehensive user study (Q2(1) Section C above) demonstrates PPC's superiority over cropping methods across multiple scenarios. To provide objective validation beyond subjective user preferences, we employ VILA to evaluate 800 image pairs comparing PPC perspective transformations (last frame) against SOTA cropping results from Gencrop (AAAI24) and Cropper (CVPR25). VILA evaluated both PPC-transformed images and crop-enhanced images on the same 0-1 aesthetic quality scale. The percentage improvements in aesthetic scores compared to original images were calculated into different ranges:

This result shows that perspective-based recomposition offers additional advantages in addressing challenges compared to cropping-based approaches.

W3. CA metric subjectivity and annotator bias.

We have conducted an inter-annotator reliability analysis. Our PQA model training employed 8 expert annotators. For the dataset, videos are independently annotated by 3 annotators using pairwise comparison across three dimensions.

To quantitatively evaluate the consistency between annotators, we calculated Fleiss' Kappa coefficient, which measures the degree of agreement among multiple raters.

We can observe that our annotations demonstrate substantial consistency across three metrics.

We sincerely thank Reviewer 5rXU for recognizing the novelty of our work, particularly appreciating our innovative approach to photography composition through video models that transcend traditional 2D cropping methods. Below, we provide detailed responses to each concern. The "W" represents weakness.

W1. About the PQA method.

We appreciate the reviewer for this point. We agree that as video generation models continue to advance, the reliance on PQA for quality assessment will naturally decrease. The improved capabilities of future video generation models will likely reduce the need for explicit quality filtering. Our technique will benefit from these continuous advancements in video generation models, making this research area promising for long-term development.

However, we think that the evaluation model like PQA still has its unique value beyond video generation models. We are actively developing our next-generation system using Unreal Engine 5-based data collection pipelines to generate higher fidelity training scenes. In this context, the evaluation model becomes particularly valuable for automated data collection workflows: it can efficiently evaluate whether a generated scene is worth collecting by providing composition quality scores, enabling large-scale automated curation of high-quality training data without manual intervention. This automated scene assessment capability will be instrumental in scaling up our data collection efforts.

W2. About the model hallucination and detail changes affecting sensitive scenarios like portrait photography.

We thank the reviewer for this concern about consistency issues. For sensitive scenarios like portrait photography, detail inconsistencies and hallucinations are inevitable. These issues primarily occur in background regions, especially when the model needs to generate background content not present in the original image. Actually, it is an inherent challenge for generative models.

However, we find that these background detail distortions do not affect our understanding of the perspective transformation guidance provided by the model. This aligns with our core objective of providing perspective transformation guidance through videos.

Additionally, to minimize the occurrence and impact of these consistency problems, we also make some efforts:

(1) Angle Limitation. During data generation, we limit the maximum angle difference (20°) between the first and last frames of training videos. In our initial experiments without this limitation, the model often exhibited severe hallucinations by generating completely new scenes. By constraining the angle difference, we ensure the model learns to perform gradual perspective transformations rather than generating entirely different scenes, which significantly reduces hallucination issues. We found this angle limitation to be an effective solution for maintaining visual consistency while still allowing meaningful perspective changes.

(2) RLHF. Even after implementing angle limitation, we still observe some videos with larger transformations that significantly affect the main subjects. Upon analyzing these problematic cases, we discover an interesting phenomenon: while these videos exhibit notable distortions, their perspective changes actually aligned well with human aesthetic preferences, but they deviated from the ground truth (GT) transformation directions. This led us to reconsider whether strict GT adherence was necessary—perhaps alignment with human preferences would be sufficient. Therefore, we introduce a two-stage RLHF training to relax the rigid GT constraints and allow the model to learn from human preference feedback rather than forcing exact GT compliance. This approach addresses the fundamental trade-off between GT fidelity and natural aesthetic transformations that humans find appealing. The effectiveness of this RLHF approach is demonstrated in Figure 6(a) in the paper: without RLHF, the cat's head directly deviates from its original position during perspective transformation, violating natural perspective relationships. After incorporating RLHF, the cat's head maintains proper perspective consistency, following realistic spatial transformations that align with human expectations.

(3) Future Improvements with Unreal Engine 5. As mentioned above, we are developing our next-generation system using Unreal Engine 5-based data generation pipelines to further enhance scene consistency. This game engine approach provides significantly more controlled environments for generating training data. With perfect lighting control, we can eliminate the lighting inconsistencies that commonly occur in real-world 3D reconstruction. The precise camera tracking capabilities allow for exact trajectory control, ensuring consistent perspective transformations. Additionally, the engine maintains perfect structural consistency across all frames while avoiding the reconstruction artifacts and hallucinations that are inherent in current image-to-video approaches. This controlled data generation environment will be instrumental in reducing consistency issues.

W3. About the overshooting.

We thank the reviewer for the careful review. Precise control over an exact "best perspective" is indeed challenging due to two main factors: (1) the subjective nature of perspective aesthetics itself, as different photographers may have varying preferences for what constitutes the "best" perspective for a given scene. (2) the inherent characteristics of video generation models.

We initially explore a more straightforward angle prediction approach, designing classification models for four-directional angle predictions (5°, 10°, 15° increments). This classification-based method shows promising results and improved accuracy with increased data.

Angle Prediction Classification Results.

However, we find that simply predicting discrete angles do not align with our original intention. We aim to provide intuitive, educational guidance that demonstrates the transformation process, which led us to persist in adopting the video generation approach for creating instructional perspective transformation videos.

About the writing.

1. Usage of the word "optimal" in subjective contexts.

We appreciate the feedback on writing clarity and will replace "optimal" with "improved" or "enhanced perspective" throughout the paper to avoid subjective claims.

2. Logical issues with sentences starting with "While".

We will revise the "While" problem and further polish our paper.

For L21, the revised version is: "Master photographers, such as those in Magnum Photos, require professional knowledge and extensive training, making quality photography expensive and challenging for ordinary people."

For L25, the revised version is: "Numerous approaches have been developed for image cropping, including saliency-based methods, learning-based techniques, and reinforcement learning strategies."

3. Missing NICER reference in the related work section.

We will add the NICER reference (Fischer et al.) to the related work section on "Evaluation with Human Performance".