CropReward: A Human-Preference Benchmark and Reward Model for Aesthetic Image Cropping
Abstract Overview
This paper studies aesthetic image cropping through the lens of human preference rather than relying only on geometric overlap or standalone aesthetic scoring. The authors introduce CropBench, a benchmark with 106,995 human preference pairs over 1,019 images, where each candidate set includes outputs from 14 cropping methods plus the original image. Using this benchmark, they show that common criteria such as IoU and off-the-shelf image aesthetic assessment models have limited alignment with human judgments. They then propose CropReward, an MLLM-based reward model that scores a crop relative to its original image, and use it to build the large-scale CropReward-300K dataset and train the cropping model CropAlign.
Novelty
The main novelty is the combination of a human-preference benchmark for aesthetic cropping and a reward model that explicitly compares a crop against the uncropped original image. The work is also distinctive in using preference-derived relative scoring to supervise both evaluation and optimization, enabling training of a cropping model without human-annotated crop boxes.
Results
On the CropBench test split, CropReward aligns more strongly with human preferences than existing evaluation criteria and fine-tuned baselines, reaching SRCC values of 0.763 on general images and 0.787 on portraits. The model also generalizes to GAICD-v2 and FGAesthetics-Crop, where it outperforms compared out-of-domain methods and even surpasses the in-domain FGAesQ on the latter benchmark. Using CropReward as supervision, CropAlign achieves the best overall scores reported on CropBench, FLMS, and FCDB, and wins 75.7% of pairwise comparisons in a user study against six competing methods.
Key Points
- CropBench provides human preference annotations that include comparisons among crops and against the original image, exposing gaps between standard AIC evaluation metrics and actual human judgments.
- CropReward is designed to score whether a crop improves or degrades the original image, and its pairwise difference learning objective improves agreement with fine-grained human preferences.
- CropReward is used not only for evaluation but also for data curation and reward-guided training, leading to the CropReward-300K dataset and the preference-aligned cropping model CropAlign.