acceptodds
Under review as a conference paper at ICLR 2027

ShutterMuse: Capture-Time Photography Guidance with MLLMs

Abstract

Computational photography has largely focused on post-capture optimization, while many photographic failures could be prevented at capture time. Capture-time assistance requires more than aesthetic scoring or crop prediction: it must determine whether the current view should be preserved, refined, or abandoned, and guide how a subject should pose within the scene. We formulate this problem as capture-time photography guidance and introduce ShutterBench, a benchmark evaluating both sides of the camera. Photographer-side guidance evaluates composition decisions (keep/refine/reject) and spatial refinement, while subject-side guidance evaluates scene-conditioned pose recommendation. Our evaluation reveals complementary limitations: general-purpose MLLMs make reasonable composition decisions but struggle with precise spatial refinement, whereas specialized cropping models localize effective crops but cannot reason over the broader decision space; MLLMs also generate valid poses, but their aesthetic quality remains limited. To enable unified learning, we construct ShutterData, comprising 130K samples with textual rationales and structured composition and pose supervision, and develop ShutterMuse, a unified MLLM trained with supervised and reinforcement fine- tuning. ShutterMuse achieves the strongest overall photographer-side performance on ShutterBench and competitive subject-side performance at substantially lower inference cost, demonstrating its potential to shift computational photography from post-hoc optimization toward actionable assistance before capture.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multimodal Large Language Models, Capture-Time Photography Guidance, Composition Decision and Refinement, Scene-Conditioned Pose Recommendation, Visual Aesthetics, Benchmarking

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_iM3oOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a benchmark, a dataset, and an MLLM-based model all designed for capture-time photography guidance covering two aspects of photography: (1) photographer-side decision making and composition refinement; and (2) subject-side pose generation and recommendation. Through extensive experiments, the authors demonstrate the benefits and superiority of their benchmark, dataset, and model.

Critical strengths

The work is presented very well, is well-motivated, and offers valid contributions. Although MLLMs have been used in the curation of the proposed dataset and benchmark, the use of experts in the loop and self-verification mechanisms makes the collected data more reliable. The model has been trained with task-specific reward models through GRPO RL training which have been demonstrated to enhance the capabilities of the model. In addition to quantitative experiments, the authors validate the superiority of the proposed model through a subjective study.

Critical weaknesses

I have only a few points here:

• (Minor) Based on my understanding of the literature, the first contribution claim is too broad. The paper claims formulation of capture-time photography guidance as its contribution, while prior works such as Venus and [1] propose capture-time aesthetic guidance and composition decision making (Keep/Refine), respectively. The authros are encouraged to narrow down this claim and reposition their novelty.

• (Minor) I’m wondering whether it is possible for the model to get stuck in a rejection loop (e.g. the photographer wants to take a photo of an object or a scene that has little aesthetic value regardless of the composition). To make the model more practical, it would be interesting to think of these cases and design a mechanism to break the loop.

• (Minor) From a practical standpoint, the “reject” composition decision should come with clear instructions as to how to resolve the composition or capturing issue, helping the photographer recapture the scene. Based on the examples shown in the paper, the textual rationale for “reject” decision only states the problem without instructions for refinement.

• (Minor) Accroding to the paper, general-purpose MLLMs and cropping methods have complementary capabilities: MLLMs are better at making composition decisions while cropping methods excel at applying refinements. Their integration, therefore, can create a good baseline: an MLLM makes the decision and if refinement is needed a cropping method is called up. The authors are encouraged to discuss this baseline or if time allows add it to their set of baselines.

[1] Li, Jiawan, et al. "Towards smart point-and-shoot photography." 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2025.

Questions for authors

I don't have any major concerns and vote for acceptance of this paper. The minor points above are my suggestions for improvement. I thank authors for putting together this interesting piece of work.

3 comments
AuthorsOpenReview ↗

Response to reviewer iM3o

We thank the reviewer for the positive assessment and constructive suggestions for improvement. We address the questions below. All changes have been incorporated into the revised manuscript and highlighted in red.

Question1:

The authors are encouraged to narrow down this claim and reposition their novelty.

Response: Thank you for your valuable feedback. We agree that prior work, including Venus [1] and SPAS [2], has already introduced capture-time aesthetic guidance. We therefore frame our first contribution as extending the formulation of capture-time photography guidance to cover both photographer-side composition decisions and refinement and subject-side pose recommendation. Fig. 13 in Appendix H of the revised manuscript illustrates this extension by comparing ShutterMuse with conventional post-capture cropping, Venus [1], and SPAS [2].

Question2:

Design a mechanism to break the reject loop.

Response: Thank you for your insightful observation. A mechanism like this could help break the rejection loop: if the same issue triggers Reject again, or if the total number of Reject decisions for the current shot exceeds a predefined threshold, the interface would offer two options: Keep adjusting or Capture anyway. Choosing the latter would end the guidance loop immediately. This gives the photographer the final say even when the model continues to suggest an adjustment.

From a training-data perspective, a complementary approach would be to construct multi-turn correction trajectories that include the previous recommendation, the updated view, and user feedback on whether the adjustment helped or was feasible. Expert-verified responses would teach the model to reassess successful adjustments, propose alternatives when an action fails, and defer to the photographer when no feasible improvement remains, rather than repeat the same instruction. This would require conditioning on interaction history and is a direction for future work.

Question3:

From a practical standpoint, the “reject” composition decision should come with clear instructions as to how to resolve the composition or capturing issue, helping the photographer recapture the scene.

Response: We totally agree with your suggestion regarding Reject decisions and have revised the annotations so that each rationale gives the photographer a concrete next step. Please see our response to Reviewer hxvz’s Weakness2 and Appendix A.1 Fig\. 9 of the revised manuscript for details. After the revision, each Reject decision comes with a concrete suggestion for improving the shot. The guidance covers camera position and orientation eg˙,˙changingcameraheight,adjustingpitch,orsteppingbacke\.g\., changing camera height, adjusting pitch, or stepping back, framing eg˙,˙excludedistractingcluttere\.g\., exclude distracting clutter, and camera settings eg˙,˙adjustingfocusorfocallengthe\.g\., adjusting focus or focal length.

Question4:

An MLLM makes the decision and if refinement is needed a cropping method is called up can create a good baseline

Response: Thank you for suggesting this two-stage baseline. We have added it to Table 1 of the revised manuscript: Gemini-3.0-Flash, which achieved the highest Macro-F1 among the baselines, makes the three-way decision. When it predicts Refine, Venus [1], which achieved the highest IoU among the baselines, provides the crop. Compared with this two-stage baseline, the revised ShutterMuse improves IoU from 66.65% to 74.74%, reduces BDE from 0.071 to 0.052, and increases [email protected] from 53.12% to 68.25%. It also achieves a higher KSR ∗∗781˙3**78\.13% vs\. 70\.83%** and MLLM-Score **0\.69 vs\. 0\.62**, with a higher Macro-F1 ∗∗700˙9**70\.09% vs\. 68\.99%**.

MethodIoU (%) ↑BDE ↓[email protected] (%) ↑RSR (%) ↑KSR (%) ↑Macro-F1 (%) ↑MLLM-Score ↑
Gemini-3.0-Flash + Venus (two-stage)66.650.07153.1274.6370.8368.990.62
ShutterMuse74.740.05268.2569.4078.1370.090.69

Thank you again for your thoughtful suggestions, which have helped us clarify our contribution and improve the guidance and baseline comparisons.


[1] Venus: Benchmarking and Empowering Multimodal Large Language Models for Aesthetic Guidance and Cropping, CVPR 2026.

[2] Towards Smart Point-and-Shoot Photography, CVPR 2025.

I thank the authors for the additional clarifications and experiments and for taking into account my recommendation on repositioning their novelty. I believe their revision has improved the quality of their work and continue to recommend clear acceptance.

Finally, I suggest that authors consider filling out the empty spaces in the main manuscript by adding back content from Appendix for the revised/final version.

Dear Reviewer iM3o,

Thank you very much for your valuable feedback and follow-up. Your suggestions have helped us better position our contribution, improve the actionable guidance, and strengthen the baseline comparisons. We will also move relevant appendix content into the available space in the main text, as suggested.

We sincerely appreciate your thoughtful review and time.

Best regards, The Authors

Reviewer_2r6gOpenReview ↗
Rating
3(+1)
Confidence
3(−1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces ShutterMuse, a multimodal large language model for capture-time photography guidance that unifies composition decision-making, composition refinement, and human pose recommendation. The authors construct ShutterData, a dataset containing 130K samples, and introduce a new benchmark, ShutterBench.

Critical strengths

1.The research problem has practical significance. The paper extends conventional image cropping to a three-way decision framework (keep/refine/reject) and incorporates scene-conditioned human pose recommendation, broadening the scope of intelligent photography assistance.

2.The dataset construction and experimental evaluation are relatively comprehensive. The authors develop a large-scale dataset with composition and pose annotations and conduct comparisons against multiple baselines. ShutterMuse achieves 74.72% IoU and 70.10% Macro-F1 on ShutterBench. The paper also includes ablation studies, cross-dataset evaluation, and expert-based subjective assessments.

Critical weaknesses

1.The central claim of capture-time photography guidance is insufficiently validated.

Although the paper emphasizes capture-time photography guidance, composition guidance is primarily formulated as crop prediction on static images, while pose recommendation relies mainly on generating 2D keypoints. Important aspects of real-world photography, such as camera positioning, shooting angles, and dynamic adjustments of subjects relative to the scene, are not explicitly modeled. Furthermore, the paper categorizes severe image tilt and exposure problems as reject cases, even though such issues can often be corrected through adjustments during capture. Consequently, there remains a substantial gap between the proposed task formulation and practical capture-time guidance, with limited validation in real-world shooting scenarios.

2.The experimental comparisons do not sufficiently establish the advantages of the unified framework.

In Table 1, specialized cropping methods such as InstructCrop and Venus do not natively support keep/reject decisions, placing them at an inherent disadvantage on the corresponding metrics. Although ShutterMuse also achieves improvements in crop localization accuracy, this raises concerns about the fairness of the overall comparison. Moreover, while the paper integrates composition guidance and pose recommendation into a single MLLM, ShutterBench evaluates these tasks separately. Joint composition and pose recommendation, as well as refinement of existing poses, are demonstrated only qualitatively. The absence of comparisons between single-task and multi-task training further makes it difficult to determine whether unified learning provides meaningful benefits.

3.The practical quality of pose recommendations remains limited.

As shown in Table 3, ShutterMuse achieves a pose aesthetics score of only 0.36, underperforming some image-editing models. Furthermore, the 2D skeleton representation cannot adequately capture depth relationships and physical interactions between the human body and the surrounding scene.

Questions for authors

How does the proposed framework distinguish capture-time photography guidance from conventional post-capture cropping?

What benefits does joint training provide over independently trained composition and pose recommendation models? Can the authors include single-task versus multi-task ablations and quantitative evaluations of the joint guidance capabilities?

How does the model ensure that its 2D pose recommendations are physically feasible and executable in real-world scenes, particularly when depth and body–environment contact are involved?

10 comments
AuthorsOpenReview ↗

Response to Reviewer 2r6g – Part 1

We sincerely thank the reviewer for recognizing the practical value of our approach. We address the questions below. All changes have been incorporated into the revised manuscript and highlighted in red.

Weakness1/Question1 : The central claim of capture-time photography guidance is insufficiently validated.

Important aspects of real-world photography, such as camera positioning, shooting angles, and dynamic adjustments of subjects relative to the scene, are not explicitly modeled.

Response : We apologize if our original explanation did not clearly convey the distinction from conventional image cropping. In particular, the wording "the current view should be preserved, refined, or abandoned" could suggest that Reject means giving up on the photograph. Our intention is instead to identify issues in the current view that cropping cannot resolve, so that the photographer can adjust the shot and have the updated view reassessed. To make this clearer, we have revised the abstract to read: "it must determine whether to preserve the current view, refine it, or retake the shot, and guide how a subject should pose within the scene."

Secondly, we appreciate your suggestion to make this guidance more actionable. We have revised the Reject annotations so that each rationale gives the photographer a concrete next step. For issues with a clear corrective action, we used templates. For example, a rationale that previously identified a tilted horizon now explicitly advises the photographer to level the camera before taking the shot. For more complex cases, six professionals annotated the appropriate shooting guidance. The revised annotations cover camera position and orientation (e.g., changing camera height, adjusting pitch, or stepping back), framing (e.g., exclude clutter), and camera settings (e.g., adjusting focus or focal length).

We retrained ShutterMuse on ShutterData with the revised Reject annotations. The table below compares the original and revised models on the photographer-side split of ShutterBench. Fig. 9 in Appendix A.1 of the revised manuscript illustrates the shooting instructions produced by the revised model.

We also evaluated the impact of retraining on the revised Reject annotations. The two tables below compare model performance before and after retraining. After retraining on the revised Reject annotations, crop localization and pose recommendation performance remained comparable to the original model.

ModelIoU (%) ↑BDE ↓[email protected] (%) ↑RSR (%) ↑KSR (%) ↑Macro-F1 (%) ↑MLLM-Score ↑
Original74.720.05368.2568.6679.1770.100.69
Revised74.740.05268.2569.4078.1370.090.69
ModelPlausibility ↑Interaction ↑Aesthetics ↑Mean ↑
Original0.800.330.360.50
Revised0.820.340.350.50

Weakness2 / Question2 : The experimental comparisons do not sufficiently establish the advantages of the unified framework.

What benefits does joint training provide over independently trained composition and pose recommendation models? Can the authors include single-task versus multi-task ablations and quantitative evaluations of the joint guidance capabilities?

Response: We agree that specialized cropping models [1, 2] do not natively support three-way composition decisions and should not be compared on decision-making metrics. In Table 1 of the revised manuscript, comparisons with these models are now limited to crop localization metrics (IoU, BDE, and [email protected]), while decision-making comparisons are restricted to methods that support the three-way task.

To evaluate the benefits of unified learning, we added composition-only and pose-only training ablations (Appendix F, Tables 9 and 10 of the revised manuscript). Compared with composition-only training, joint training yields slight improvements in refinement precision. For decision-making, ShutterMuse exhibits a trade-off: RSR decreases from 80.60% to 69.40%, while KSR increases from 52.08% to 78.13%. Overall, Macro-F1 across the three decision classes increases from 68.99% to 70.09%, indicating a modest improvement in overall decision-making despite the decline in reject accuracy. Compared with pose-only training, it produces a larger improvement in pose aesthetics (0.35 vs. 0.30).

Training setupIoU (%) ↑BDE ↓[email protected] (%) ↑RSR (%) ↑KSR (%) ↑F1 (%) ↑MLLMScore ↑Plausibility ↑Interaction ↑Aesthetics ↑Mean ↑
Composition only74.280.05365.5880.6052.0868.990.66————
Pose only———————0.820.330.300.48
ShutterMuse (joint)74.740.05268.2569.4078.1370.090.690.820.340.350.50
AuthorsOpenReview ↗

Response to Reviewer 2r6g – Part 2

Response to Weakness2 / Question2 (continued): We also evaluated ShutterMuse on simultaneous composition and pose recommendation using a held-out set of 500 joint-task examples, comparing it with a Qwen-Image-Edit + Venus (cascade) baseline. The baseline first uses Qwen-Image-Edit [3] to recommend a pose, then applies Venus [2] to the resulting image to recommend a composition, and finally evaluates the resulting pose and composition outputs. The table below, also included as Table 11 in Appendix G of the revised manuscript, shows that ShutterMuse achieves better results across all six metrics while producing both outputs in one joint prediction.

MethodIoU (%) ↑BDE ↓[email protected] (%) ↑Plausibility ↑Interaction ↑Aesthetics ↑
Qwen-Image-Edit + Venus (cascade)42.140.1621.400.710.240.28
ShutterMuse (one joint prediction)67.000.06739.000.820.300.33

Weakness3 / Question3 : The practical quality of pose recommendations remains limited.

ShutterMuse achieves a pose aesthetics score of only 0.36, underperforming some image-editing models. Furthermore, the 2D skeleton representation cannot adequately capture depth relationships and physical interactions between the human body and the surrounding scene.

Response : Firstly, we would like to clarify that ShutterMuse’s approach to pose recommendation reflects a trade-off between inference cost and pose quality. Image-editing models recommend a pose by rendering the subject in the scene, whereas ShutterMuse directly predicts a 2D skeleton. In Table 3 of the revised manuscript, ShutterMuse produces a recommendation in 4.96 seconds on average, compared with 102.61 seconds for GPT-Image-2. Its pose aesthetics score of 0.35 is the highest among the evaluated VLMs and is close to GPT-Image-2’s score of 0.38. To further evaluate pose aesthetics against image-editing methods, we added Qwen-Image-Edit-2512 [3] (20B) as an additional baseline. ShutterMuse outperforms Qwen-Image-Edit-2512 [3] across all three pose-quality with an average latency of only 4.96 seconds, a 16.6× speedup.

Fig. 3 of the revised manuscript visualizes the trade-off between Mean pose-quality score and inference time for all nine methods evaluated on subject-side ShutterBench. ShutterMuse lies on the empirical Pareto frontier, achieving a Mean score of 0.50 at 4.96 seconds per recommendation, the lowest measured latency among the evaluated methods.

MethodPlausibility ↑Interaction ↑Aesthetics ↑Mean ↑Time (s) ↓
Image-editing models
Qwen-Image-Edit-25120.720.280.310.4482.24
Vision-language models
ShutterMuse (Ours)0.820.340.350.504.96

We also appreciate the reviewer’s concern that ShutterMuse may not adequately capture depth relationships and physical interactions between the subject and the scene. ShutterMuse alleviates your concern in two ways. First, as described in Appendix A.2 (“Rationale generation and review”), ShutterMuse generates a step-by-step rationale to help users understand how the subject interacts with the surrounding scene in the 2D skeleton visualization. Second, as described in Appendix A.2 (Pose Representation and Verification), ShutterMuse predicts a visibility field that specifies the visibility state of each keypoint. Our VLM-as-a-judge evaluation assesses the correctness of these predictions as part of the Plausibility criterion. ShutterMuse achieves a score of 0.82 on this dimension (Table 3), providing evidence that it can capture depth relationships expressed through occlusion.


Thank you again for your constructive feedback, which has helped us clarify our task formulation and strengthen the experimental evaluation.


[1] InstructCrop: Teaching Multimodal Large Language Models to Crop Aesthetic Images, ACM MM 2025.

[2] Venus: Benchmarking and Empowering Multimodal Large Language Models for Aesthetic Guidance and Cropping, CVPR 2026.

[3] Qwen-Image Technical Report, arXiv 2025.

AuthorsOpenReview ↗

Friendly Reminder: Rebuttal Period for ICLR submission

We kindly remind the reviewer that we have addressed your concerns in our rebuttal. We clarified the distinction between capture-time guidance and conventional post-capture cropping, and revised Table 1 so that specialized cropping models are compared only on applicable crop-localization metrics.

To further evaluate the unified framework, we added single-task versus joint-training ablations in Tables 9–10, as well as a quantitative comparison of simultaneous composition and pose recommendation against a Qwen-Image-Edit + Venus cascade baseline in Table 11.

We also added Qwen-Image-Edit-2512 as an additional pose-generation baseline and discussed the limitations of 2D skeleton representations for modeling depth and body–scene contact.

We sincerely hope these clarifications and additional experiments help alleviate your concerns. Please kindly let us know if you have any further questions; we would be happy to provide additional clarification.

Once again, thank you very much for your valuable feedback and review!

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

The paper proposes using an MLLM for two tasks: (1) composition guidance, which decides whether to keep, refine, or reject an image and predicts an appropriate crop, and (2) scene-conditioned human pose recommendation, represented as 2D keypoints. To achieve this goal, the authors construct ShutterData, a dataset comprising 130K composition and pose examples, along with 10K additional examples for joint composition and pose guidance. They then fine-tune Qwen3-VL-8B using supervised fine-tuning (SFT), followed by GRPO reinforcement fine-tuning for composition guidance, with rewards based on decision correctness, bounding-box IoU, and salient-content preservation. The authors also introduce ShutterBench to evaluate both tasks. Experiments demonstrate improvements over general-purpose MLLMs and specialized cropping models for composition guidance, while achieving competitive pose recommendation quality compared with image-editing models at substantially lower inference cost

Critical strengths
  • The paper provide useful benchmark and dataset for photography guidance.
  • Strong empirical performance with relatively comprehensive evaluation.
Critical weaknesses
  • The proposed photographer-side guidance has limited action space (keep/refine/reject), with refinement represented solely by a cropping bounding box. Practical guidance should provide more informative recommendation such as moving the camera left/right/up/down, changing the viewpoint, adjusting the camera angle. When an image is classified as "reject," the model provides no structured guidance on how to improve the next capture. Thus, the proposed task formulation is much closer to conventional image cropping and quality assessment than to the broader capture-time photography guidance claimed by the paper.
  • With more detailed instruction as above, it's more challenging for synthesizing as frontier-LLM model could not accurately judge/filter correctly.
  • Are shutterdata and shutterbench constructed from similar pipeline/source? if yes, it raise the question about performance transfers to real capture-time photography.
Questions for authors
  • The figure 5 demonstrate the improvement when having larger data (with gemini filter), it would be informative if the author can compare the performance when increasing dataset size with and without gemini (and other MLLM model) filter. Since performance can be increase with noisy + larger data, this can help demonstrate how effective/important the filter model is.
7 comments
AuthorsOpenReview ↗

Response to Reviewer hxvz – Part 1

We sincerely thanks for your valuable comments and thoughtful feedback, which have helped us further improve and clarify our work. We address the questions below. All changes have been incorporated into the revised manuscript and highlighted in red.

Weakness1: Unclear differences from Conventional Image Cropping

The proposed task formulation is much closer to conventional image cropping and quality assessment than to the broader capture-time photography guidance claimed by the paper.

Response: We apologize if our original explanation did not clearly convey the distinction from conventional image cropping. In particular, the wording "the current view should be preserved, refined, or abandoned" could suggest that Reject means giving up on the photograph. Our intention is instead to identify issues in the current view that cropping cannot resolve, so that the photographer can adjust the shot and have the updated view reassessed. To make this clearer, we have revised the abstract to read: "it must determine whether to preserve the current view, refine it, or retake the shot, and guide how a subject should pose within the scene."

Conventional image cropping starts with a captured photo whose content is fixed and typically assumes that a crop is needed. ShutterMuse extends this setting with Reject, which identifies issues requiring a change to the shot eg˙,˙levelingaseverelytiltedcamerae\.g\., leveling a severely tilted camera, and Keep, which indicates that the current framing can be retained. It also recommends how a subject should pose within the scene before capture. Following the reviewer's suggestion, our revised Reject guidance provides concrete instructions for adjusting the shot (see our response to Weakness2 below).

To further clarify these distinctions, Fig. 13 in Appendix H of the revised manuscript compares the decision workflows of post-hoc cropping eg˙,˙InstructCrop[1]e\.g\., InstructCrop [1], Venus [2], SPAS [3], and ShutterMuse. It illustrates how ShutterMuse selects among Reject, Refine, and Keep for each new view, with reassessment after user adjustments rather than abandonment of the photograph.


Weakness2: Lack of more Specific Photography Guidance

Practical guidance should provide more informative recommendation such as moving the camera left/right/up/down, changing the viewpoint, adjusting the camera angle. When an image is classified as "reject," the model provides no structured guidance on how to improve the next capture.

Response: Thank you for your suggestion on improving Reject. In the original version, ShutterMuse flagged issues that needed to be addressed before a crop could improve the composition, helping users prepare for Refine. Following your suggestion, we have revised the Reject annotations so that the rationale gives the photographer a concrete next step. Specifically, we used templates for issues with a clear corrective action. For example, a rationale that previously identified a tilted horizon now instructs the photographer to level the camera before taking the shot. For more complex cases, six professionals annotated the appropriate shooting guidance. The revised annotations cover camera position and orientation eg˙,˙changingcameraheight,adjustingpitch,orsteppingbacke\.g\., changing camera height, adjusting pitch, or stepping back, framing eg˙,˙recomposingtoexcludedistractingcluttere\.g\., recomposing to exclude distracting clutter, and camera settings eg˙,˙adjustingfocusorfocallengthe\.g\., adjusting focus or focal length.

We retrained ShutterMuse on ShutterData with the revised Reject annotations. Fig. 9 in Appendix A.1 of the revised manuscript shows examples of the concrete shooting instructions produced by the revised model.

We also evaluated the impact of retraining on the revised Reject annotations. The two tables below compare model performance before and after retraining. After retraining on the revised Reject annotations, crop localization and pose recommendation performance remained comparable to the original model.

ModelIoU (%) ↑BDE ↓[email protected] (%) ↑RSR (%) ↑KSR (%) ↑Macro-F1 (%) ↑MLLM-Score ↑
Original74.720.05368.2568.6679.1770.100.69
Revised74.740.05268.2569.4078.1370.090.69
ModelPlausibility ↑Interaction ↑Aesthetics ↑Mean ↑
Original0.800.330.360.50
Revised0.820.340.350.50

Further details of the revised Reject annotations and shooting guidance are provided in Appendix A.1 of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer hxvz – Part 2

Weakness3: Unclear sources and distributions of the Training and Evaluation Data

Are shutterdata and shutterbench constructed from similar pipeline/source? if yes, it raise the question about performance transfers to real capture-time photography.

Response: ShutterBench was constructed primarily through human annotation, using a different pipeline from the one used to build ShutterData.

The training and evaluation images were collected from the same platforms but partitioned into non-overlapping sets before annotation.

On the photographer side, ShutterData starts with 12K expert-annotated images and scales to 100K through EMDP pseudo-labeling Sec\. 3\.1, Fig\. 6(d). For ShutterBench, experts annotated all 567 evaluation images following the seed-set annotation protocol. Each of the 337 Refine test images has 3–5 expert reference boxes, allowing several valid compositions, whereas each Refine training example has one target box Sec\. 3\.2 and Appendix A\.1. ShutterMuse also achieves the highest IoU and lowest BDE among the evaluated methods on the external FLMS benchmark [4], providing further evidence that its framing predictions generalize beyond ShutterBench.

On the subject side, we constructed ShutterData using the pipeline shown in Fig. 10 in Appendix A.2. ShutterBench includes 100 scenes originally photographed without people to evaluate pose recommendation under real-world conditions Sec\. 3\.2.

Fig. 11 in Appendix A.3 summarizes the data distributions of ShutterData and ShutterBench. By using distinct construction procedures and a more balanced distribution of evaluation examples, we designed ShutterBench to better reflect model performance in real-world capture-time photography.


Question1: Insufficient filter analysis.

it would be informative if the author can compare the performance when increasing dataset size with and without gemini (and other MLLM model) filter.

Response: We agree that evaluating the filtering step is crucial to validating our EMDP pipeline. To address your concern, we conducted an additional ablation study comparing three data construction settings. We used Gemini-3.0-Pro as the filter, used GPT-5.5 as the filter, or omitted filtering altogether. Across all three settings, we held both the number of EMDP iterations and the final dataset size constant. The table below, also included as Table 5 in the revised manuscript, reports the resulting model performance. EMDP achieves the best results with Gemini-3.0-Pro as the filter. These results support filtering as an effective component of EMDP.

EMDP filterIoU (%) ↑BDE ↓[email protected] (%) ↑RSR (%) ↑KSR (%) ↑Macro-F1 (%) ↑MLLM-Score ↑
No filter71.010.05764.6946.2754.1764.310.52
GPT-5.574.720.05465.2861.9473.9666.50.62
Gemini-3.0-Pro74.740.05268.2569.4078.1370.090.69

Thank you once again for your valuable feedback and insightful questions. We hope our responses have clarified our approach and addressed your concerns. We would appreciate any additional insights you may have and look forward to the possibility of further discussion.


[1] InstructCrop: Teaching Multimodal Large Language Models to Crop Aesthetic Images, ACM MM 2025.

[2] Venus: Benchmarking and Empowering Multimodal Large Language Models for Aesthetic Guidance and Cropping, CVPR 2026.

[3] Towards Smart Point-and-Shoot Photography, CVPR 2025.

[4] Automatic Image Cropping using Visual Composition, Boundary Simplicity and Content Preservation Models, ACM MM 2014.

AuthorsOpenReview ↗

Friendly Reminder: Rebuttal Period for ICLR Submission

We kindly remind the reviewer that we have addressed your suggestions in our rebuttal. In particular, we clarified how ShutterMuse differs from conventional post-capture cropping, revised the Reject annotations to provide concrete shooting instructions, and added examples of the revised guidance in Fig. 9. We also clarified the relationship between ShutterData and ShutterBench, and added a controlled filter ablation comparing no filtering, GPT-5.5, and Gemini-3.0-Pro.

We sincerely hope these revisions and additional experiments help alleviate your concerns. Please kindly let us know if you have any further questions; we would be happy to provide additional clarification.

Once again, thank you very much for your valuable feedback and review!

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal

Dear PCs, SACs, ACs, and Reviewers,

We sincerely appreciate the time, effort, and constructive feedback devoted to reviewing our submission. In response to the reviewers’ comments, we have carefully revised the manuscript, added new experiments, clarified the task formulation and data construction process, and expanded the discussion of related work. All reviewers have confirmed that their questions have been resolved. We hope the following summary will help clarify the changes made in the revised version.


Addressing the Concerns Raised by Reviewer hxvz

  • Distinction from conventional cropping: We clarified the capture-time keep/refine/reject formulation and revised the abstract and workflow description.
  • Specific Reject guidance: We revised the reject annotations to provide concrete instructions for camera adjustment, framing, and camera settings.
  • Training and evaluation data: We clarified the data sources, construction procedures, partitioning, and distributions of ShutterData and ShutterBench.
  • Filter ablation: We compared two MLLM filters and no filtering under fixed settings of three EMDP rounds and 100K samples, supporting the benefits of filtering.

Addressing the Concerns Raised by Reviewer 2r6g

  • Fair baseline comparisons: We restricted comparisons with specialized cropping models to crop-localization metrics.
  • Two-stage baseline: We added a Gemini-3.0-Flash + Venus baseline, with ShutterMuse achieving better framing accuracy, overall decision Macro-F1, and MLLM-Score.
  • Pose quality and efficiency: We added Qwen-Image-Edit-2512 as a baseline and demonstrated ShutterMuse’s competitive pose quality at lower latency.
  • Benefits of unified training: Single-task and refine-only ablations support the benefits of joint training and three-way decisions for pose aesthetics.
  • Related work: We expanded the discussion of PhotoFramer, PhotoBot, and InstaPose and clarified their differences from ShutterMuse.

Addressing the Concerns Raised by Reviewer iM3o

  • Contribution positioning: We clarified how ShutterMuse extends capture-time guidance to cover both photographer-side composition and subject-side pose recommendation.
  • Reject-loop handling: We discussed user-controlled stopping options and future training with multi-turn feedback to address repeated rejection.
  • Concrete Reject instructions: We revised the reject rationales to provide actionable shooting suggestions.
  • Baseline and workflow comparisons: We added the Gemini + Venus baseline and a workflow comparison with conventional cropping, Venus, and SPAS.
  • Use of main-text space: We will incorporate relevant appendix content into the available space in the main text, as suggested.

We sincerely thank all reviewers for their constructive feedback, which has helped us improve the clarity, evaluation, and presentation of our work.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.