Work

CABRIA: contrastive attribute preference bridge for personalized image aesthetic assessment

Year: 2026
Type: article
Abstract: Personalized Image Aesthetic Assessment (PIAA) aims to predict an individual user’s aesthetic ratings from a small set of images scored by that user. Generic aesthetic predictors are able to provide useful population-level knowledge, but they collapse the user-specific deviations that define personal taste. Direct few-shot adaptation can also entangle shared image quality with sparse user evidence, making it difficult to identify what should be personalized. To address this problem, we introduce CABRIA , a support-residual-conditioned framework that casts personalization as a user-specific correction over a global aesthetic prior. CABRIA first learns the generic prior and contrastively pretrains the representation used for support retrieval, then converts support-rating residuals into positive and negative preference memory. A Cross-Attention Bridge provides a structured route that connects positive and negative support evidence with query attribute tokens, and per-user gradient adaptation estimates the final residual correction. Under the 37-user FLICKR-AES protocol, CABRIA with ResNet-101 achieves SROCC scores of 0.694 and 0.760 for 10-shot and 100-shot personalization, respectively. A strict user-disjoint control that excludes validation- and test-user ratings from Stage 1 shows a statistically significant 10-shot improvement over the frozen prior and all matched adaptation baselines, while the multiplicity-corrected 100-shot comparisons do not establish an advantage for the complete pipeline. The code is available at CABRIA’s GitHub repository .
Language: English
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7235898818
FWCI:
Cites: 41
Cited by:
Related to: 10