acceptodds
Withdrawn from ICLR 2027

PhotoGuider: Composition-Aware Cropping and 360° View Framing with Explainable Photographic Assessments

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Traditional image cropping assumes the photograph already contains the composition worth keeping, and asks only where to crop. A full 360 scene offers no such guarantee: the viewing direction in a scene has to be discovered before anything can be framed. We present PhotoGuider, a measure-then-frame method that answers both questions with compact learned scorers over one shared set of composition-aware features. The designed features are extracted using foundation and task-specific vision models and structured into four photographic tokens: Subject, Composition, Aesthetics, and Integrity. Fine-tuned on and prompted with these measured tokens, our MLLM explainer provides composition-related explanations grounded in measurement rather than image content only. To our knowledge, PhotoGuider is the first method to combine explicit composition-based direction selection in 360 panoramas with refined framing, camera-parameter output for native re-rendering, and measurement-conditioned explanations. On two standard 2D cropping benchmarks, PhotoGuider achieves the highest average mIoU performance against ground-truth crops. In a blind ranking study, 120 participants ranked three methods over 100 panoramic scenes from the public Poly Haven HDRI library, placing our framings first in 53.2% of cases, against 27.3% and 19.5% for two strong alternatives.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
cropping, composition, aesthetic, panoramaa

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_FpeKOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents PhotoGuider, a composition-aware photography framework that integrates image cropping, 360° view selection, framing refinement, and interpretable photographic explanations. The method achieves competitive results on standard image cropping benchmarks and demonstrates promising performance in a user study on panoramic scenes.

Critical strengths

1.Practical research relevance. This paper extends conventional image cropping to 360° panoramic scenes, integrating view selection, composition optimization, and photographic explanations into a unified framework.

2.Strong experimental performance. PhotoGuider achieves competitive average performance on the FCDB and FLMS image cropping benchmarks. For 360° panoramic framing, the authors conduct a subjective evaluation involving 120 participants and 100 scenes, achieving a first-place preference rate of 53.2%.

3.Interpretability of composition assessment. The paper organizes photographic assessment into four feature groups: Subject, Composition, Integrity, and Aesthetics, which are used for composition scoring, cropping refinement, and natural language explanations. Compared with approaches that rely entirely on black-box models, this design provides more interpretable evidence for framing decisions.

Critical weaknesses

1.Insufficient distinction from prior work.

The proposed method primarily integrates existing techniques, including visual feature extraction, composition assessment, candidate ranking, cropping refinement, and large language model-based explanations, into a unified framework. However, most of its core technical components are based on established methods, and the paper does not clearly introduce a fundamentally new algorithmic contribution. This makes the repeated claims of being the "first" somewhat confusing.

For example, CACNet (Hong et al., CVPR 2021) already explicitly models photographic composition rules through Key Composition Maps and uses composition information to guide image cropping. Furthermore, Towards Smart Point-and-Shoot Photography (Li et al., CVPR 2025) has already investigated composition assessment and camera pose adjustment using 360° panoramic data, enabling composition guidance during photography. The authors also do not include an experimental comparison with Li et al. (CVPR 2025), even though that work has not been open-sourced.

Although the paper cites these studies, it does not sufficiently clarify the substantive methodological differences. The four Photographic Tokens mainly organize and combine features extracted from pretrained models, and their methodological novelty is not adequately justified. As a result, the proposed approach appears closer to a system-level integration of existing techniques than to a clearly established independent algorithmic contribution.

2.Insufficient experimental comparisons for 360° panoramic framing.

The paper emphasizes 360° view selection as a central contribution, yet the corresponding experiments only compare against CACNet and Venus. Both methods were originally designed for conventional image cropping rather than view selection in panoramic photography. The authors use the composition classification confidence of CACNet as its panoramic view score. However, classification confidence does not necessarily correspond to overall photographic composition quality, raising concerns about whether this adaptation adequately represents the baseline's capabilities.

The paper does not provide a direct comparison or sufficiently detailed experimental analysis against the panoramic composition guidance method proposed by Li et al. (CVPR 2025). Consequently, the current experiments do not fully establish the advantages of PhotoGuider for panoramic photography guidance.

Moreover, the ablation studies mainly focus on conventional 2D image cropping, with limited analysis of the panoramic view selection module. It remains unclear how much each component, including the four feature groups, the view scoring model, and the candidate view search strategy, contributes to the final performance.

3.Lack of computational efficiency and practical deployment evaluation.

Photography composition guidance requires not only high-quality framing recommendations but also reasonable inference latency, computational efficiency, and deployment feasibility, particularly in practical scenarios such as mobile photography.

The proposed framework relies on several pretrained vision models, including DINOv3, SAM, and CLIP, and evaluates 120 candidate viewing directions during panoramic framing. The full mode additionally involves large language model inference, potentially introducing substantial computational overhead.

Although Appendix A.5 reports that the lightweight decision layer runs within milliseconds on a CPU, this does not reflect the inference speed of the complete system. The paper does not report key metrics such as end-to-end inference latency, per-module runtime, GPU memory consumption, or overall hardware requirements. It also provides no evaluation of deployment performance on mobile devices.

The current experiments primarily demonstrate framing quality on offline panoramic images, without establishing whether the system can achieve the response speed required for practical photography guidance. Therefore, composition quality metrics and subjective evaluation results alone are insufficient to demonstrate the usability of the proposed approach in real-world photography scenarios.

4.Insufficient evaluation of the practical effectiveness of photographic explanations.

The paper mainly evaluates explanation quality using the contradiction rate between generated explanations and the supplied measurements. However, this metric only reflects consistency with the underlying numerical features. It does not adequately establish whether the explanations are consistent with photographic composition principles or whether they provide meaningful guidance to photographers.

As one of the paper's emphasized contributions, the explanation module lacks user evaluations assessing explanation accuracy, comprehensibility, and practical usefulness.

Questions for authors

Methodological novelty: Compared with CACNet (Hong et al., CVPR 2021) and Towards Smart Point-and-Shoot Photography (Li et al., CVPR 2025), what are the substantive algorithmic innovations of the proposed approach in composition modeling and panoramic view selection? Beyond organizing existing visual features, what capabilities do the four Photographic Tokens provide that cannot be achieved by previous methods?

Experimental comparisons: Why does the panoramic evaluation only include CACNet and Venus as baselines, without comparing against Li et al. (CVPR 2025)? Could the authors design an appropriate comparison protocol that accounts for differences in the input and output formulations of the two approaches? Additionally, could the authors provide further ablation studies specifically for panoramic view selection to demonstrate the independent contributions of different components?

Computational efficiency: What are the end-to-end inference latency, peak GPU memory consumption, and hardware requirements of the complete system when processing a single 360° panorama? How do the computational costs of the Fast and Full modes differ in practice?

Practical applicability: Can the proposed method run on standard mobile devices or support near-real-time photographic composition guidance? If high-performance GPUs or cloud-based inference are required, can the resulting interaction latency satisfy the requirements of practical photography?

I am willing to raise my rating if the authors provide convincing explanations or evidence addressing the concerns above. I also welcome counterarguments if the authors believe any of these concerns are unfounded, and I am open to revising my assessment accordingly.

Reviewer_Sh2SOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces PhotoGuider, a unified framework that integrates 360° panoramic view selection, automatic image cropping, and explainable photographic evaluation. The method leverages four categories of photographic features and combines a lightweight ranking model, a composition fine-tuning module, and a fine-tuned large multimodal model to generate cropped images along with corresponding explanations. Experimental results on two standard image-cropping datasets show strong performance.

Critical strengths
  1. The decisions of the system is transparent by defining photographic quality through four interpretable dimension.
  2. The authors compare the approach against several existing methods using two standard datasets for image cropping, and further validate its effectiveness through a user study conducted in 360° scenes. The experimental evaluation is comprehensive, and the results demonstrate a competitive level of performance.
Critical weaknesses
  1. The paper provides limited discussion of concrete real-world applications and user needs for the proposed system.
  2. The overall framework is constructed by integrating existing vision foundation models, hand-crafted photographic features, LightGBM, and MLLMs. While the system design is comprehensive, the core methodological innovation is relatively limited.
  3. For two cropping windows that are highly similar in terms of spatial position and size, one would expect the composition scores to exhibit a degree of local stability, with the corresponding explanations remaining semantically consistent. However, the paper does not analyze how composition scores fluctuate under perturbations such as translation or scaling, nor does it verify the consistency of explanations across adjacent cropping windows.
  4. The paper lacks an analysis of the factors driving performance gains in the explanation module. It remains unclear whether improvements are from data filtering, fine-tuning strategies, or the model architecture itself. Furthermore, the absence of comparisons against more robust multimodal backbones makes it difficult to assess the necessity and effectiveness of the current fine-tuning approach.
Questions for authors

Refer to Critical Weaknesses.

Reviewer_17yXOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper builds PhotoGuider, a system for 360° panorama auto-framing and image cropping. It mimics human photographer workflow, defines four interpretable photographic tokens (Subject, Composition, Integrity, Aesthetics). The pipeline first picks camera viewing direction on 360 scene, then ranks & refines crop boxes with LightGBM ranker/gates, and uses fine-tuned MLLM to generate measurement-grounded explanation text. They test on standard 2D cropping benchmarks and run human user study for 360 panorama scenes, getting decent quantitative and human preference results.

Critical strengths

1.The overall pipeline logic makes sense. Authors nicely imitate real-world photographer step-by-step workflow: scan scene, pick view direction, adjust frame, output final photo. This human-aligned thinking is well reflected in system design. 2. Experimental results look solid and support their claims. They report standard metrics on two public 2D cropping datasets, do ablation study for four tokens, and run large-scale human user study with 120 participants for the 360 framing task. 3.The four photographic tokens are intuitive. They separate different photographic principles and use them for direction selection, crop ranking and explanation generation together.

Critical weaknesses

1.The problem scope is narrow. The 360-degree view framing task is a tiny area with limited audience. To me this is mostly engineering adaptation for a very specific application scenario, lacks broader scientific research value. 2.Limited generalizable insight for broader ML communities. It is largely an application stacking existing foundation models together. There is little core methodological novelty that can inspire or transfer to other research domains. Most components are off-the-shelf models assembled for this photography task. 3.I have little experience on the view framing task, but I think it is more suitable for applied conference, not ICLR.

Questions for authors

No questions.

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

Related papers

Discussion (0)

Sign in to comment.