PaperlayerView exact revision on arXiv

Source-linked AI summary

CropReward: A Human-Preference Benchmark and Reward Model for Aesthetic Image Cropping

Xiangfei Sheng, Xiaofeng Pan, Leida Li, Qian Zhang, Biao Wang, Johan Wagemans, Weisi Lin

arXiv:2610.34221v2cs.CV

TL;DR

AIC evaluation commonly relies on geometry metrics and isolated aesthetic scores, which do not fully reflect human preferences between crops and originals or among similar crops. The paper builds CropBench and CropReward, then uses CropReward to curate training data and optimize CropAlign. CropReward aligns more closely with human preferences than compared criteria, while CropAlign achieves state-of-the-art evaluation performance and is preferred in a user study.

  • Problem

    Existing AIC criteria show limited alignment with human preferences, and their ability to assess crop-versus-original preference and subtle differences among crops remains unclear.

    Source: §I · pp. 1–2

  • Method

    The paper builds CropBench with pairwise crop preferences and ties, trains CropReward on these preferences, and uses its scores to curate data and guide CropAlign.

    Source: §I · p. 2, §III · p. 4

  • Results

    CropReward aligns more closely with human preferences than compared criteria and generalizes to external benchmarks; CropAlign achieves state-of-the-art performance and is preferred over specialized models and MLLMs in a user study.

    Source: §VII · p. 13

  • Takeaways & Limitations

    CropReward provides an evaluation and optimization signal based on human crop preferences, and CropAlign can be trained without human annotation in its training process.

    Source: §VII · p. 13

Abstract

from arXiv · show

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 AIC existing criteria. CropAlign also surpasses the state-of-the-art AIC methods, demonstrating the advantage of CropReward in guiding AIC model optimization.

I. INTRODUCTION

AIC needs evaluation that reflects whether crops improve on the original and distinguishes subtle preferences, but geometry metrics and isolated aesthetic scores do not directly capture both. The paper introduces CropBench and CropReward, then uses CropReward to build a dataset and train CropAlign.

I. INTRODUCTION

  • Existing geometry-based metrics rely on expert reference boxes, while isolated IAA scores do not account for removed content or changes relative to the original.Different pleasing crops can depart from a reference box, and high crop aesthetics do not necessarily mean people prefer the crop to the original.Source: §I · pp. 1–2
  • Reliable AIC evaluation should assess preferences between crops and originals and distinguish subtle differences among candidate crops, but existing criteria's fit to these requirements remains unclear.CropReward is trained on CropBench preference annotations; CropAlign uses its rewards for reinforcement-learning optimization.Source: §I · p. 2
  • CropBench contains 106,995 human preference pairs across outputs from 14 AIC methods and original images, enabling systematic evaluation of criteria and methods.The benchmark includes explicit ties and 1,019 images, as described in the introduction.Source: §I · p. 2
  • CropReward is an MLLM reward model that compares each crop with its original image and predicts aesthetic improvement or degradation, aligning better with human preferences than existing criteria.Its scores are relative to the original, addressing both what cropping removes and fine-grained differences between crops.Source: §I · p. 2
  • CropAlign reaches state-of-the-art performance on CropBench, FLMS, and FCDB after training on CropReward-300K curated without human annotation.The dataset contains over 300,000 original images paired with CropReward-selected crop boxes.Source: §I · p. 2

II. RELATED WORK

Prior AIC work uses crop-box regression, candidate ranking, and aesthetic or composition assessment, while datasets and protocols vary in their annotations and evaluation targets. These approaches do not directly establish whether people prefer a crop to its original image.

A. Aesthetic Image Cropping

  • Crop regression predicts a target box, whereas candidate ranking scores candidate crops; neither comparison alone establishes preference over the uncropped original.Regression penalizes deviations from a target box even when alternatives are equally good; ranking can model human crop preferences when the original is included.Source: A. Aesthetic Image Cropping · p. 3

B. AIC Datasets and Evaluation Protocols

  • AIC datasets span expert crop boxes, candidate-quality judgments, synthetic crops, and visually similar-image preferences, with evaluation protocols tied to those annotation types.IoU and boundary displacement measure box alignment, while ranking metrics compare predicted scores with mean opinion scores; these do not directly assess crop-versus-original preference.Source: B. AIC Datasets and Evaluation Protocols · pp. 3–4

C. Image Aesthetic and Composition Assessment

The section motivates evaluating crops against human preferences, since existing aesthetic and composition scores do not directly show whether a crop is preferred to its original. It introduces CropBench’s pairwise annotations, CropReward’s crop-to-original comparison, and the benchmark’s image-selection scope.

C. Image Aesthetic and Composition Assessment

  • IAA methods predict aesthetic scores or rating distributions, while ICA models assess spatial arrangement; existing IAA/ICA models have limited discrimination and absolute scores do not establish preference over the original, [9].Prior work includes multi-scale representations, vision-language aesthetic representations, natural-language explanations, and relative rankings,,,, [28],,; composition work models salient regions and patterns [30] and explores explanations and improvement guidance, [32].Source: C. Image Aesthetic and Composition… · p. 4

D. Reward Models for Visual Preference Alignment

  • CropReward learns from human preferences among candidate crops and crop-original pairs, then gives each crop a positive or negative score relative to the original.The score can be used as an evaluation criterion and as a reward for AIC model optimization.Source: D. Reward Models for Visual Preference… · p. 4

III. CROPBENCH: A HUMAN-PREFERENCE BENCHMARK FOR AIC

  • CropBench annotates candidate pairs with an explicit tie option, enabling comparisons among crops and between crops and their original uncropped images.Candidate sets include outputs from AIC models as well as the original image.Source: §III · p. 4

A. Image Collection

  • 50.4% of the initial general pool and 55.0% of the portrait pool were removed during screening; the final benchmark contains 510 general and 509 portrait images.Images were selected for potential improvement through cropping, screened by an MLLM and three aesthetic experts, then sampled across aesthetic levels, orientations, and scene or subject categories.Source: A. Image Collection · p. 4

B. Candidate Crop Generation

CropBench pairs model-generated crops with original images, collects human preferences using an MLLM-guided annotation process, and converts those labels into per-image preference scores. Its construction covers three stages, from candidate generation through tie-aware BT-score calculation.

B. Candidate Crop Generation

  • Ten specialized AIC models and four general-purpose MLLMs each contribute one crop per image; with the original, this yields 15 candidates and 106,995 comparison pairs.The pool spans regression-, ranking-, RL-, and MLLM-based AIC methods.Source: B. Candidate Crop Generation · p. 4

C. Hybrid Preference Annotation

  • Five MLLM judges determine how many human annotators assess each pair, while humans provide the final preference labels.Pairs receive two, four, or six annotators according to judge agreement; unresolved disagreements receive another review round.Source: C. Hybrid Preference Annotation · pp. 4–5

D. BT-score Calculation

  • A tie-aware Davidson extension of Bradley–Terry estimates each image’s candidate scores from its annotated outcomes, anchoring the original image at zero.Positive and negative scores indicate improvement and degradation relative to the original, respectively.Source: D. BT-score Calculation · pp. 5–6

E. Benchmark Analysis

CropBench comparisons show that AIC methods vary in preference scores and cropping extent. GPT-5.5 leads the evaluated methods, but cropping can still reduce preference, and retained area alone does not indicate crop quality.

E. Benchmark Analysis

  • GPT-5.5 achieves the highest mean BT score for both general and portrait images, exceeding all ten evaluated specialized AIC models under zero-shot evaluation.Venus has the second-highest mean score, and performance varies substantially across MLLMs.Source: E. Benchmark Analysis · p. 6
  • Several methods have negative mean BT scores, and even the highest-ranked methods produce some crops with negative scores.This supports assessing crops relative to their original images, including for methods that perform well on average.Source: E. Benchmark Analysis · p. 6
  • A2-RL retains 86.9% of general-image area and 86.0% of portrait area, while Qwen3.5-4B retains 25.4% and 23.6%, respectively.GPT-5.5 scores higher than both while retaining approximately 72% of image area in each category; retained area alone is not informative about crop quality.Source: E. Benchmark Analysis · p. 6

IV. CROPREWARD: A HUMAN-ALIGNED REWARD MODEL FOR AIC

CropReward scores a candidate crop relative to its original image and learns from differences in human-derived preference scores. Its architecture uses an MLLM scoring head, while pairwise difference learning supervises score differences across candidate pairs.

A. Formulation

  • CropReward learns a scoring function that takes an original image and candidate crop as input and predicts scores whose differences match human-derived BT preference-score differences.Training compares candidate crops with one another and with the original image; expert-annotated reference boxes are not required.Source: A. Formulation · pp. 6–7

B. Model Architecture

  • The model compares each crop with its original image, then uses the final-layer representation of a learnable <AES> token and a two-layer MLP to predict its score.The visual patch merger, rank-16 LoRA adapters, <AES> embedding, and MLP head are trained, totaling 62M trainable parameters.Source: B. Model Architecture · p. 7

C. Pairwise Difference Learning

  • The NPD loss matches predicted and BT score differences across all candidate pairs, including equivalent-preference pairs, rather than supervising preference order alone.CropReward is trained jointly on general and portrait images from the CropBench training split.Source: C. Pairwise Difference Learning · p. 7

D. Inference

CropReward scores crops relative to the original image, and its scores are used to select pseudo-label crops from unlabeled images for CropReward-300K.

D. Inference

  • A crop’s inference score is computed relative to the original image, with positive scores favoring the crop and negative scores favoring the original.Subtracting the original-image score preserves the ranking of candidates from the same image.Source: D. Inference · p. 7

A. Construction of CropReward-300K Dataset

  • CropReward scores crops from eight top-ranked AIC models per domain on 517,889 Unsplash images and selects the highest-scoring crop per image as a pseudo-label.Images are ranked by selected-crop score, and the top 60% in each domain are retained.Source: A. Construction of CropReward-300K… · pp. 7–8
  • CropReward-300K contains 310,735 images: 149,879 general images and 160,856 portraits, each paired with its selected crop, score, and generating method.The selected crops are used as training data without additional human annotation.Source: §V · p. 7, A. Construction of CropReward-300K… · p. 8

B. Reward-Guided Cropping

CropAlign is initialized by supervised fine-tuning and then optimized with group-relative policy optimization using CropReward scores. The experiments assess reward alignment, generalization, ablations, and its effectiveness for model optimization.

B. Reward-Guided Cropping

  • CropAlign is initialized by supervised fine-tuning Qwen3.5-4B on CropReward-300K, then optimized with GRPO while frozen CropReward evaluates generated crops.The policy generates a textual rationale followed by a crop box; Fig. 8 depicts the reward-guided update.Source: B. Reward-Guided Cropping · p. 8, Figure 8 · p. 8
  • GRPO samples groups of responses, extracts valid crop boxes, and uses CropReward’s relative crop scores to construct rewards and group-normalized advantages.Rewards map the relative score through a sigmoid and include a format bonus; the RL stage uses no target crop boxes or additional human annotation.Source: B. Reward-Guided Cropping · p. 8

VI. EXPERIMENTS

  • The experiments assess human-preference alignment, generalization, training-objective and input-protocol ablations, and CropReward’s effectiveness in guiding AIC optimization.The reward-guided cropping workflow is illustrated in Fig. 8.Source: §VI · p. 8, Figure 8 · p. 8

A. Experimental Setup

The experiments evaluate preference alignment, generalization, and cropping performance on CropBench and external AIC datasets, using correlation, accuracy, and expert-crop comparisons. Implementation details specify training schedules and model settings.

A. Experimental Setup

  • CropBench evaluation uses 200 test images and reports agreement with human preferences; GAICD-v2 and FGAesthetics-Crop test generalization, while FLMS and FCDB evaluate cropping models.The external sets include 500 GAICD-v2 test images, 168 FGAesthetics-Crop image series, 500 FLMS images, and 345 valid FCDB test images.Source: A. Experimental Setup · p. 8
  • CropBench agreement is measured with PLCC, SRCC, pairwise accuracy, and top-1/top-3 accuracy; FGAesthetics-Crop uses series-level selection accuracy and rank correlation.CropBench scores are averaged over images within each domain; its pairwise accuracy evaluates 105 annotated pairs per image.Source: A. Experimental Setup · pp. 8–9
  • CropReward is trained for three epochs with AdamW, a 2.5×10−5 learning rate, and a cosine decay schedule after 3% warmup.CropAlign receives two SFT epochs followed by one GRPO epoch with group size 4; runs use two NVIDIA RTX PRO 6000 Blackwell GPUs.Source: A. Experimental Setup · p. 9

B. Alignment with Human Preferences

CropReward aligns with human cropping preferences more closely than the evaluated criteria, including when compared with expert-box-based and fine-tuned baselines. Qualitative examples also show disagreements between existing criteria and annotator rankings.

B. Alignment with Human Preferences

  • IoUmean outperforms all evaluated off-the-shelf models on every metric for both image categories, but agreement with expert boxes does not fully reflect human preferences.Multiple well-composed crops can differ from reference boxes, and IoU requires expert-annotated boxes.Source: B. Alignment with Human Preferences · pp. 9–10
  • Off-the-shelf models’ top-1 accuracy reaches at most 0.150 on general images and 0.130 on portraits, while fine-tuning three evaluators improves them but does not surpass CropReward.Among off-the-shelf models, FGAesQ leads IAA models on PLCC, SRCC, and pairwise accuracy; EditReward leads reward models on most metrics.Source: B. Alignment with Human Preferences · p. 10
  • CropReward achieves SRCC of 0.763 on general images and 0.787 on portraits, exceeding IoUmean by 0.140 and 0.130, respectively, without expert crop-box references.It also exceeds the strongest fine-tuned baseline, the GAIC evaluator, by 0.135 and 0.170 in SRCC, respectively; pairwise accuracy is 0.761 and 0.777, and top-3 accuracy is 0.600 and 0.590.Source: B. Alignment with Human Preferences · p. 10
  • Qualitative examples show existing criteria misranking crops against annotators, while CropReward assigns negative scores to all four crops in a portrait example where annotators prefer the original.IoUmean also ranks a visually similar portrait crop above another despite annotators preferring the latter.Source: B. Alignment with Human Preferences · pp. 10–11, Figure 9 · p. 10

C. Generalization to External Benchmarks

CropReward generalizes from CropBench to external benchmarks with different candidate crops and annotation types, outperforming compared models. Ablations favor the NPD objective and retaining the original image as input.

C. Generalization to External Benchmarks

  • On GAICD-v2, CropReward reaches PLCC 0.766 and SRCC 0.716, versus 0.413 and 0.386 for ArtiMuse; on FGAesthetics-Crop, it surpasses in-domain FGAesQ on all four metrics without fine-tuning.The external benchmarks use grid-anchor candidates with MOS and pairwise preferences among visually similar crops, respectively.Source: C. Generalization to External Benchmarks · pp. 11–12

D. Ablation Studies

  • The NPD objective achieves the best results on nine of ten metrics across general and portrait domains, outperforming score regression and pairwise ranking losses.NPD matches predicted score differences to Bradley–Terry preference-score differences over all candidate pairs.Source: D. Ablation Studies · p. 12
  • Removing the original image input lowers all ten metrics; SRCC drops by 0.156 and 0.163, while general-image and portrait ACCtop1 fall from 0.300 to 0.180 and 0.320 to 0.130.Source: D. Ablation Studies · p. 12

E. Cropping Performance of CropAlign

CropAlign performs strongly under existing evaluation criteria and is preferred in a user study. The paper also reports interpretable crop rationales and uses CropReward supervision to develop an AIC model without human-annotated crop boxes.

E. Cropping Performance of CropAlign

  • CropAlign achieves the best Overall score on CropBench, FLMS, and FCDB; GRPO further improves Overall on all three, with CropBench IoU rising from 0.675 to 0.698.CropAlign-SFT is already competitive with the strongest baselines, and CropReward scores increase after optimization.Source: E. Cropping Performance of CropAlign · p. 12
  • 75.7% of 3,000 user-study comparisons favor CropAlign over six AIC methods, with all six comparisons significant at p < 0.001.Win rates range from 73.2% against GPT-5.5 to 78.6% against InstructCrop.Source: E. Cropping Performance of CropAlign · pp. 12–13
  • CropAlign produces textual rationales consistent with its crop boxes without human-written rationale supervision during SFT or GRPO.Examples describe reducing excess sky and foreground while preserving subjects or salient scene content.Source: E. Cropping Performance of CropAlign · p. 13
  • In qualitative examples, CropAlign preserves the statue and its base while removing unnecessary surroundings, and frames an entrance more tightly than GPT-5.5.The tighter crop retains the entrance and people while emphasizing surrounding activity.Source: E. Cropping Performance of CropAlign · p. 13

VII. CONCLUSION

  • The authors use CropReward to build CropReward-300K and train CropAlign without human annotation; they report state-of-the-art performance under existing criteria.Source: §VII · p. 13

Related papers

Loading 2610.34221v2…