Open World Image Aesthetic Assessment

Mingxiang Liao, Tianren Ma, Xijin Zhang; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Findings, 2026, pp. 9791-9801

Prevailing Image Aesthetic Assessment (IAA) models assume a single, generic aesthetic criterion, yet real-world aesthetics are fundamentally pluralistic (diverse themes, unique criteria) and open (new themes constantly emerging). To bridge this gap, we propose the Open-World IAA (OW-IAA) task, requiring models to exhibit both versatility across diverse seen themes and generalization to unseen ones. To address this, we propose Induce-and-Adapt, a novel framework that first induces a generalizable reasoning policy via jointly optimizing for score prediction and reasoning and then efficiently adapts to new themes by optimizing policy against crowd-simulated preferences--achieved without human annotation. To evaluate this task, we introduce OA-Bench, the first benchmark for OW-IAA, which reveals current methods are ill-equipped for this challenging setting. Our method builds new state-of-the-art on both OA-Bench and TAD66K, and achieves 0.102 PLCC improvement over prior best. It also demonstrates a synergistic lift by improving generalization to unseen themes without sacrificing versatility on seen ones.