acceptodds
Withdrawn from ICLR 2027

Discrete Annotation, Continuous Preference: Rethinking Supervision for Accurate and Generalizable Aesthetic Image Cropping

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Aesthetic image cropping aims to identify the optimal crop of an image in terms of aesthetics and composition. While supervision based on annotated data is fundamental, the field has been hindered by a long-standing problem: existing datasets suffer from (1) human subjectivity and (2) rigid discreteness confined to fixed sampling grids. These flawed annotations not only limit the accuracy and generalization of trained models but also severely distort fair evaluation. To overcome this, we propose to model human cropping preference as a multi-peaked, continuous, and sharp field over the crop space. We introduce the Continuous Preference Field (CPF), which recovers a dense preference landscape from discrete annotations through (1) peak clustering, (2) off-lattice refinement, (3) negative shaping, and (4) field assembly. Based on this, we train CPIC, a VLM-based cropping model optimized via GRPO with the CPF reward, which overcomes template collapse, achieving state-of-the-art performance and exceptional out-of-domain generalization. Finally, to resolve the long-standing benchmark evaluation crisis, we introduce CPICD, a comprehensive recalibration of existing ground-truth boxes. By leveraging the CPF to correct grid-bound artifacts across mainstream benchmarks, CPICD establishes a rigorous and reliable foundation for future cropping research. Extensive experiments and user studies demonstrate the superiority of our CPF, CPIC, and CPICD. Code will be released soon.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Aesthetic Image Cropping, Vision-Language Models, 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_u1F7OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates the limitations of discrete annotations for aesthetic image cropping. It proposes a Continuous Preference Field (CPF), which combines peak clustering, off-lattice refinement, structured negatives, and field assembly to construct a reward over crop coordinates. CPF is used to train a vision-language cropping model, CPIC, through supervised fine-tuning followed by GRPO. The same refinement procedure produces recalibrated benchmark annotations, termed CPICD. Validation includes comparisons on GAIC, FLMS, FCDB, and CPICD, component ablations, qualitative examples, and two human preference studies.

Critical strengths
  1. A thoughtful, task-specific reward design. CPF explicitly incorporates multiple acceptable compositions, tolerance to local coordinate changes, and penalties for undesirable boundary changes. Combining these properties provides a plausible alternative to optimizing geometric overlap alone. The ability to reuse existing annotations makes this a practically relevant direction, and the reward construction is the paper’s strongest technical contribution.

  2. Broad experimental coverage. The paper evaluates several families of cropping methods and provides substantial ablations of supervision strategies, field components, and refinement scores. The human studies additionally examine both predicted crops and recalibrated annotations. This breadth helps assess individual design choices and provides useful empirical support beyond selected qualitative examples.

Critical weaknesses
  1. Insufficient novelty positioning and incomplete comparisons. GAIC [1] already discusses non-unique crops, local redundancy, and the limitations of IoU. Prior work also addresses diverse compositions [2] and annotation bias causing quasi-trivial predictions [3]. These established observations are insufficiently distinguished from CPF’s specific contribution. Several relevant methods [2–6] are omitted, weakening the novelty and state-of-the-art claims.

  2. The incremental benefit of the complete reward design is not sufficiently established. Results are mixed: CPIC improves FCDB IoU but trails Venus on FLMS IoU. In Table 6, refined targets with an IoU reward already match full CPF’s output-box count and FLMS LAION score; CPF adds only 0.007/0.013 IoU on FLMS/FCDB. Without repeated-run statistics, the reliability of these additional gains remains unclear. The evidence therefore falls short of the strong claims about CPF’s necessity and consistently superior generalization.

  3. The evaluation protocol and evidence for resolving template collapse are inadequate. Appendix A describes predicting one box per image, whereas standard GAIC SRCC requires scores over multiple candidate crops. The conversion to candidate rankings and the treatment of off-grid predictions for ACC₁₀ are unspecified. Furthermore, counting distinct boxes across different images measures coordinate repetition rather than meaningful compositional diversity. These issues limit both reproducibility and the interpretation of the reported improvements.

[1] Hui Zeng, Lida Li, Zisheng Cao, and Lei Zhang. Reliable and Efficient Image Cropping: A Grid Anchor Based Approach. CVPR, 2019.

[2] Gengyun Jia, Huaibo Huang, Chaoyou Fu, and Ran He. Rethinking Image Cropping: Exploring Diverse Compositions From Global Views. CVPR, 2022.

[3] Zhiyu Pan, Yinpeng Chen, Jiale Zhang, Hao Lu, Zhiguo Cao, and Weicai Zhong. Find Beauty in the Rare: Contrastive Composition Feature Clustering for Nontrivial Cropping Box Regression. AAAI, 2023.

[4] Zhiyu Pan, Zhiguo Cao, Kewei Wang, Hao Lu, and Weicai Zhong. TransView: Inside, Outside, and Across the Cropping View Boundaries. ICCV, 2021.

[5] Chao Wang, Li Niu, Bo Zhang, and Liqing Zhang. Image Cropping With Spatial-Aware Feature and Rank Consistency. CVPR, 2023.

[6] Yukun Su, Yiwen Cao, Jingliang Deng, Fengyun Rao, and Qingyao Wu. Spatial-Semantic Collaborative Cropping for User Generated Content. AAAI, 2024.

Questions for authors

Check the weakness.

Reviewer_RQvLOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes a CPF that combines refined crop anchors and structured negatives to model aesthetic preferences, uses CPF rewards to train a VLM cropper via GRPO and recalibrate benchmark annotations, and evaluates these contributions through benchmark comparisons, ablation studies, and human preference studies. This is an interesting research direction, and image composition is an important task.

Critical strengths
  1. The motivation is clear, and the proposed approach is easy to understand.
  2. Incorporating negative examples into learning for image composition offers a useful perspective.
  3. The paper includes extensive experiments, and the figures and tables are clearly presented.
Critical weaknesses
  1. There is already substantial work on identifying better candidate compositions by shifting, scaling, or adjusting the aspect ratio of a crop around the subject. Examples include: Deep Cropping via Attention Box Prediction and Aesthetics Assessment, Learning the Relation between Interested Objects and Aesthetic Region for Image Cropping, Focusing on Your Subject: Deep Subject-Aware Image Composition Recommendation Networks, etc...However, the authors introduce a relatively new aspect by incorporating negative examples into learning to help distinguish good compositions from problematic ones.

  2. For the composition task considered in this paper, adjusting an initial crop produces candidates with the same subject but slightly different compositions. The key difficulty is that GAIC, SAMP-Net, and YOLOv8 alone may struggle to reliably assess how these subtle compositional differences affect aesthetics. For example, subject retention measured using YOLOv8 bounding boxes may not be sufficiently sensitive to partially cropped hands or feet. This criterion is also less suitable for half-body portraits or scenes containing multiple subjects. The GAIC scorer itself learns from existing annotations and can therefore inherit their errors. Even for human annotators, distinguishing the aesthetic quality of such similar compositions can sometimes be extremely difficult. Therefore, the paper provides insufficient insight into the most critical issue: how to build an evaluation or reward model that distinguishes subtle aesthetic differences. The fused score or criteria may themselves contain substantial noise. Yet the authors emphasize “reducing reliance on subjective annotations and biases inherited by one scorer” while overlooking errors in the scoring criteria themselves and whether their combination can reliably distinguish subtle compositional differences.

  3. Improvements in several cropping metrics after training with CPF do not establish that the method adequately captures the broad claim that “Aesthetic cropping preference is a multi-peaked, continuous, and sharp distribution.” Improving downstream cropping metrics and accurately characterizing human aesthetic cropping preferences are two different problems. The existing results are not even sufficient to establish CPF's effectiveness in continuous composition assessment: can it reliably evaluate unannotated, off-grid crops with only subtle compositional differences in a way that agrees with human aesthetic judgments?

  4. The paper emphasizes that a small boundary shift can damage the subject while barely changing IoU. Specifically, it states: “Conversely, a small boundary shift can cut through a subject's head or another visually critical region while barely changing IoU.” Under this argument, crops with high overlap but poor aesthetic quality should be particularly important negative examples. However, the appendix states: “Negatives with IoU ≥ 0.85 to their associated refined peak are removed to avoid conflicting supervision.”On the one hand, the paper emphasizes that highly overlapping crops can have poor aesthetic quality because of slight subject truncation. On the other hand, it explicitly removes negatives that overlap highly with their associated refined peaks. This clearly contradicts the stated motivation. Could this rule filter out precisely those hard negatives that crop away only a small part of a hand, foot, or head? The authors need to explain this point.

  5. No demo or code is provided in the supplementary materials to enable verification of the proposed method.

Questions for authors

Please refer to Weaknesses.

Reviewer_US1gOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses the limitations of discrete cropping annotations through a Continuous Preference Field (CPF), used for reinforcement learning of a vision-language cropper and benchmark annotation refinement. Results show improved cropping quality and generalization, supported by human preference studies.

Critical strengths
  1. I find the focus on supervision limitations meaningful. The idea may also be relevant to other vision tasks involving subjective annotations.
  2. Multiple preference peaks, continuity, and sensitivity to critical cropping errors capture useful aspects of human cropping preferences.
Critical weaknesses

see questions

Questions for authors
  1. Sometimes, could the best crop be no crop at all? I wonder whether the model preserves an already well-composed image rather than adjusting its boundaries simply because it was asked to crop. A few such examples would be interesting.
  2. Can the model appreciate two different compositions of the same photo?
  3. How does the model balance subject preservation with bold framing? Portraits sometimes intentionally crop the top of a head or part of the body. Could the preservation rule favor safe but less expressive compositions?
  4. Are there examples where the metrics get worse but people prefer the crop? Given the paper’s concerns about existing annotations and evaluation metrics, such examples could be particularly informative about what actually improves.

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

Related papers

Discussion (0)

Sign in to comment.