Dual Cross-Attention for Artwork Aesthetic Assessment with Style-Conditioned Fusion and Cross-Dataset Quantile Alignment

We study artwork aesthetic assessment with optional comments and heterogeneous rating scales using dual cross-attention, style-conditioned fusion, and quantile-aligned ordinal learning. An explicit visual path handles absent comments, while corpus-specific readouts convert a shared percentile prediction to native scores. Across three training seeds, the joint model achieves mean Spearman correlations of 0.8362 on the retained APDDv2 subset and 0.7164 on AADB, compared with 0.8216 and 0.6917 without pairwise consistency. These descriptive differences do not isolate the ranking objective because tuning histories and pair budgets are not verified as matched. In a separate fixed-checkpoint analysis of 374 text-present artworks, 100 comment derangements reduce correlation from 0.9020 to a mean of 0.6449, with assignment standard deviation 0.0269, while mean textual routing weight remains almost unchanged. An added MMLQ adaptation achieves 0.8041 ± 0.0097 over five training seeds under different encoder and training settings. Historical cross-corpus diagnostics show small, comparator-dependent ordering differences whose conditional intervals include zero. Reporting labels for the main-result models were used for hyperparameter tuning, not supervised parameter fitting; consequently, the evaluation is not independent of development. The evidence distinguishes missing-text handling from misleading-text detection and supports configuration-level comparisons rather than isolated component gains.

Authors

Institutions

Publication Details

Journal
Applied Sciences
Published
2026-11-19
DOI
https://doi.org/10.3390/app162211470
Primary Topic
Image and Video Quality Assessment
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Dual Cross-Attention for Artwork Aesthetic Assessment with Style-Conditioned Fusion and Cross-Dataset Quantile Alignment

Kije Kwon, GUOWEN ZHU, Dayu Wu
Applied Sciences
Image and Video Quality Assessment
article

Dual Cross-Attention for Artwork Aesthetic Assessment with Style-Conditioned Fusion and Cross-Dataset Quantile Alignment

Kije Kwon, GUOWEN ZHU, Dayu Wu
article en

Abstract

We study artwork aesthetic assessment with optional comments and heterogeneous rating scales using dual cross-attention, style-conditioned fusion, and quantile-aligned ordinal learning. An explicit visual path handles absent comments, while corpus-specific readouts convert a shared percentile prediction to native scores. Across three training seeds, the joint model achieves mean Spearman correlations of 0.8362 on the retained APDDv2 subset and 0.7164 on AADB, compared with 0.8216 and 0.6917 without pairwise consistency. These descriptive differences do not isolate the ranking objective because tuning histories and pair budgets are not verified as matched. In a separate fixed-checkpoint analysis of 374 text-present artworks, 100 comment derangements reduce correlation from 0.9020 to a mean of 0.6449, with assignment standard deviation 0.0269, while mean textual routing weight remains almost unchanged. An added MMLQ adaptation achieves 0.8041 ± 0.0097 over five training seeds under different encoder and training settings. Historical cross-corpus diagnostics show small, comparator-dependent ordering differences whose conditional intervals include zero. Reporting labels for the main-result models were used for hyperparameter tuning, not supervised parameter fitting; consequently, the evaluation is not independent of development. The evidence distinguishes missing-text handling from misleading-text detection and supports configuration-level comparisons rather than isolated component gains.

Applied SciencesVol. 16(22)
Yeungnam University (KR)
Image and Video Quality Assessment
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon