CropReward: A Human-Preference Benchmark and Reward Model for Aesthetic Image Cropping
Authors: Xiangfei Sheng, Xiaofeng Pan, Leida Li, Qian Zhang, Biao Wang, Johan Wagemans, Weisi Lin
Organizations: School of Artificial Intelligence, Xidian University, Xi’an 710071, China · Huawei Technologies Co., Ltd., Shenzhen 518129, China · Laboratory of Experimental Psychology, Department of Brain and Cognition, KU Leuven, 3000 Leuven, Belgium · College of Computing and Data Science, Nanyang Technological University, Singapore
Aesthetic image cropping (AIC) aims to improve image composition and aesthetic quality by removing extraneous or distracting content. Over the past decade, AIC methods have evolved from specialized vision models to multimodal large language models(MLLMs). Reliable evaluation that reflects human preferences is essential for benchmarking AIC methods and guiding model optimization. However, the current AIC evaluation relies largely on geometry-based metrics such as intersection over union (IoU), as well as image aesthetic assessment(IAA) models. Although inconsistencies with human preferences have long been observed, the robustness and reliability of these evaluation criteria remain underexplored. To bridge this gap, we first construct CropBench, a large-scale benchmark comprising 106,995 human preference pairs for 1,019 images, with each candidate set containing the original image and candidate crops produced by 14 AIC methods. A systematic study on CropBench reveals that the existing criteria, ranging from IoU to off-the-shelf IAA assessment models, only exhibit limited alignment with human preferences. Building upon CropBench, we further propose CropReward, an MLLM-based reward model that evaluates a candidate crop by comparing with the original uncropped image and outputs a score indicating explicit aesthetic improvement or degradation relative to the original. In addition, we employ CropReward to construct CropReward-300k, the largest AIC dataset to date, and further develop CropAlign, a preference-aligned AIC model trained solely with CropReward supervision, without human-annotated crop boxes. Experimental results demonstrate that CropReward aligns with human preferences significantly better than existing criteria. CropAlign also surpasses the state-of-the-art AIC methods, demonstrating the advantage of CropReward in guiding AIC model optimization.
Figures & tables
Fig. 1: Existing evaluation criteria versus the proposed CropReward. (a) Geometry-based metrics compare the coordinates of predicted and expert reference boxes, while image aesthetic assessment (IAA) models score each crop in isolation. (b) CropReward takes the original image and a crop as input and outputs a score relative to the original image, with positive and negative values indicating predicted improvement and degradation, respectively. Bottom: An example from CropBench. IoU mean is computed over the three expert reference boxes shown on the left. Numbers in parentheses denote ranks.
Fig. 2: Three training paradigms for AIC. (a) Crop regression learns to predict a target crop box. (b) Candidate ranking learns to score candidate crops. (c) Reward-guided exploration uses CropReward to evaluate each generated crop by comparing with the uncropped original image. The resulting reward is used to optimize the cropping policy toward human preferences without relying on human-annotated crop boxes.
Dataset
Venue
Images
Candidate views
Annotation signal
Candidate source
Expert crops
Original image
Explicit pairs
Domain
Expert-crop datasets
ICDB [ 3 ]
CVPR 2013
950
2,850
3 expert crops/image
Expert crops
✓
✗
✗
General
FLMS [ 11 ]
ACM MM 2014
500
5,000
10 expert crops/image
Expert crops
✓
✗
✗
General
FCDB [ 12 ]
WACV 2017
1,743
–
34,130 pairwise prefs.
Expert crops + random subviews
✓
✗
✓
General
Candidate-scoring datasets
CPC [ 13 ]
CVPR 2018
10,800
259,200
Per-annotator candidate labels (0–4)
Algorithm-generated + random views
✗
✗
✗
General
TABLE I: Comparison of CropBench with representative AIC datasets. ✓: included; ✗: unavailable. Images denote the number of source images. “Original image” indicates whether the original image is included as a candidate in the comparisons, and “Explicit pairs” indicates whether candidate-pair labels are included in the public release.
Fig. 3: Four-stage construction pipeline of CropBench: image collection, candidate crop generation, preference annotation, and BT-score calculation. Each image has 14 model-generated crops and the original uncropped image as a candidate crop set. Agreement among five MLLM judges determines the number of human annotators assigned to each pair. BT scores are calculated from the final human preference labels, with the score of the original uncropped image anchored at zero.
Fig. 5: Performance and cropping extent of the 14 AIC models on CropBench. (a) Distributions of per-image BT scores for general (blue) and portrait (purple) images, with methods ordered by their overall mean scores. Diamonds denote mean BT scores, and the dashed line marks the original image score of zero. (b) Mean retained-area ratio, defined as the crop area divided by the original image area.
Fig. 6: Framework of CropReward. (a) Candidate scoring: an MLLM takes the original image and one candidate crop as input. A regression head predicts a score from the final-layer representation of the <AES> token. (b) Pairwise difference learning: the model is trained to match predicted score differences to BT preference score differences over all candidate pairs, including comparisons with the original image.
Fig. 7: Mean CropReward scores of the eight AIC models used to generate candidates in each domain, computed on the corresponding general or portrait images used to construct CropReward-300K. Scores are computed relative to the original image.
Fig. 8: Reward-guided cropping with CropReward. The LoRA-adapted policy generates a group of G responses containing a textual rationale followed by a crop box. The frozen CropReward scores each valid crop relative to the original image. These scores are used to construct rewards, which are normalized within each group to compute advantages for the GRPO update.
TABLE II: Agreement with human cropping preferences on the CropBench test split. For each candidate, IoU is computed against three expert crop boxes. The subscripts min , mean , and max indicate the minimum, mean, and maximum of these three values, respectively. Bold indicates the best result within each group in each column.
Method
GAICD-v2 [ 2 ]
FGAesthetics-Crop [ 18 ]
PLCC
SRCC
ACC
F1
s\mbox−ACC
s\mbox−SRCC
NIMA [ 24 ]
0.232
0.210
0.658
0.653
0.303
0.323
MUSIQ [ 25 ]
0.184
0.164
0.753
0.753
0.406
0.499
VILA [ 10 ]
0.351
0.332
0.733
0.733
0.406
0.458
AesCLIP [ 27 ]
0.295
0.273
0.679
0.674
0.325
0.374
ArtiMuse [ 29 ]
0.413
0.386
0.682
0.682
0.354
0.387
TABLE III: Generalization to two external benchmarks, GAICD-v2 [ 2 ] and FGAesthetics-Crop [ 18 ] , with all models applied without adaptation. PLCC and SRCC are computed within each image and macro-averaged; ACC and F1 are pair-level, whereas the s -prefixed metrics are series-level. Grey cells mark in-domain results. Bold and underline indicate the best and second-best generalization results in each column.
Fig. 9: Qualitative comparison of predicted and human preference rankings on CropBench test images from (a) the portrait domain and (b) the general domain.
Original
Lscore
Lrank
LNPD
General
Portrait
PLCC
SRCC
ACCpair
ACCtop1
ACCtop3
PLCC
SRCC
ACCpair
ACCtop1
ACCtop3
✓
✓
0.824
0.754
0.755
0.260
0.560
0.854
0.758
0.764
0.250
0.600
✓
✓
0.832
0.757
0.755
0.280
0.570
0.851
0.777
0.770
0.260
0.570
✓
0.715
0.607
0.706
0.180
0.450
0.778
0.624
0.710
0.130
0.450
✓
✓
0.833
0.763
0.761
0.300
0.600
0.865
0.787
0.777
0.320
0.590
TABLE IV: Ablation on the training objective and the input protocol of CropReward. All variants use Qwen3.5-4B and are trained for three epochs under identical settings. Original indicates whether the original image accompanies the candidate crop as input.
Fig. 10: Comparison of explanations for cropping decisions. Green and red highlight selected crop-related phrases from CropAlign and GPT-5.5, respectively.
Comparison
Win (%)
Tie (%)
Lose (%)
vs. GPT-5.5 [ 42 ]
73.2
1.4
25.4
vs. Venus [ 7 ]
74.2
0.6
25.2
vs. HCIC [ 22 ]
75.6
0.6
23.8
vs. GAIC [ 2 ]
76.2
0.6
23.2
vs. CACNet [ 4 ]
76.4
0.2
23.4
vs. InstructCrop [ 6 ]
78.6
0.2
21.2
TABLE VI: User study comparing CropAlign with six competing methods.
Fig. 11: Qualitative comparison of cropping results.
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, model, and data are available at https://github.com/zzqingz/CPIC.
Ziqing Zhang, Xiao Liu, Kai Liu +4
Shanghai Jiao Tong University · Institute of Media Technology and Experience Design, Huawei
Aesthetic image cropping aims to enhance the aesthetic quality of an image by improving its composition through spatial cropping. Previous methods often rely on saliency prediction or retrieval augmentation, ignoring the task's core requirement: a deep understanding of composition and aesthetics. Consequently, saliency-based methods struggle to make compositional trade-offs in complex scenes, while retrieval-based methods blindly refer to similar cases, lacking adaptive reasoning for unique scenes. Both approaches fail to align their automated cropping results with those of human experts. To address the above issues, we propose a novel paradigm that reformulates aesthetic cropping as a multimodal reasoning task, aiming to activate the VLM's analytical and comprehension capabilities in aesthetics. We design a Compositional Reasoning and Optimizing Preference method (CROP) that directs the VLM to think like a professional photographer. It deconstructs a complex and subjective aesthetic problem into an "analysis-proposal-decision" process, reasoning step by step through the analysis of scene elements and compositional principles. Meanwhile, our expert preference alignment module makes the model's decision consistent with human expert aesthetics. Extensive experiments across multiple datasets validate our method's superiority and component effectiveness.
Zhitong Dong, Chao Li, Jie Yu +1
Southeast University · Key Laboratory of New Generation Artificial Intelligence Technology · Alibaba Group
Explainable aesthetic image cropping requires not only localizing a visually pleasing crop but also explaining why it is preferred. Existing crop-and-explain methods largely treat explanation as post-hoc text generation and overlook composition, a key aesthetic factor that links crop decisions with interpretable reasoning. In this paper, we reformulate explainable aesthetic image cropping as a structured crop-composition-explanation problem. To support this setting, we introduce COMEX, a new benchmark built through image expansion and an IO-reversal pipeline. COMEX contains 33,161 quadruples, each consisting of an expanded image, a crop box, a composition category, and a composition-grounded explanation, enabling joint learning of crop localization, composition understanding, and explanation generation. We further propose a two-stage SFT+GRPO framework, where supervised fine-tuning establishes the structured output protocol and basic cropping ability, and GRPO further improves crop quality, composition prediction, and explanation faithfulness. We benchmark 15 large vision-language models and existing cropping methods on COMEX, establishing a comprehensive testbed for composition-grounded explainable aesthetic cropping. Experiments on both COMEX and prior benchmarks demonstrate the effectiveness and transferability of our framework, with strong performance across evaluation metrics.
Rui Yang, Wei Zhou, Dingyong Gou +5
State Key Laboratory of Mobile Network and Mobile Multimedia Technology, ZTE Corporation · School of Data Science and Institute of Artificial Intelligence, Chang’an University