From Concepts to Judgments: Interpretable Image Aesthetic Assessment

Xiao-Chang Liu; Johan Wagemans

cs.CVMar 18, 2026

From Concepts to Judgments: Interpretable Image Aesthetic Assessment

Authors: Xiao-Chang Liu, Johan Wagemans

Organizations: Research Unit Brain and Cognition, KU Leuven Leuven, Belgium

Abstract

Image aesthetic assessment (IAA) aims to predict the aesthetic quality of images as perceived by humans. While recent IAA models achieve strong predictive performance, they offer little insight into the factors driving their predictions. Yet for users, understanding why an image is considered pleasing or not is as valuable as the score itself, motivating growing interest in interpretability within IAA. When humans evaluate aesthetics, they naturally rely on high-level cues to justify their judgments. Motivated by this observation, we propose an interpretable IAA framework grounded in human-understandable aesthetic concepts. We learn these concepts in an accessible manner, constructing a subspace that forms the foundation of an inherently interpretable model. To capture nuanced influences on aesthetic perception beyond explicit concepts, we introduce a simple yet effective residual predictor. Experiments on photographic and artistic datasets demonstrate that our method achieves competitive predictive performance while offering transparent, human-understandable aesthetic judgments.

Figures & tables

Figure 1: Methods overview. We predict the aesthetic score of an image using human-understandable concepts and an inherently interpretable sparse linear model. (a) Aesthetic Concept Learning: Given an aesthetic concept $s$ ( e.g., rule of thirds) and user-defined positive/negative image sets, we learn its Concept Activation Vector (CAV) $\mathbf{c}_{s}\in\mathbb{R}^{d}$ by training a linear Support Vector Machine (SVM) to distinguish between image embeddings from the two sets. The learned CAV is orthogonal to the decision boundary. (b) Concept Subspace Construction: We aggregate the CAVs of $N_{c}$ aesthetic concepts to form a concept subspace $\mathbf{C}\in\mathbb{R}^{d\times N_{c}}$ . (c) Interpretable Aesthetic Assessment: For a given image, we extract its embedding using a pre-trained image encoder and project it onto the concept subspace. The resulting concept projection is used by an inherently interpretable sparse linear model to predict the aesthetic score. To account for nuanced influences on aesthetic judgment beyond explicit concepts, we add a residual predictor that complements the interpretable core.

Methods	SRCC	PLCC
NIMA [ 85 ]	0.708	0.711
MLSP [ 21 ]	0.725	0.726
RGNet [ 46 ]	0.710	-
PA-IAA [ 42 ]	0.720	0.728
MUSIQ [ 35 ]	0.706	0.712
DINOv2-small [ 66 ]	0.682	0.695

Table 1: Performance comparison on the AADB dataset. ‘-’ indicates an unreported metric in the original paper. Best and second-best results are highlighted in bold and underlined, respectively.

Figure 2: Learned aesthetic concept weights on the AADB dataset, ordered by importance (corresponding to

\mathbf{w}

in Eq. 4 ). The bias term is 0.538 (corresponding to

b

in Eq. 4 ).

Figure 3: Aesthetic score prediction on the AADB test image. Bottom left shows the ground truth (GT), our interpretable prediction (Interp. pred.), and hybrid prediction (Hybrid pred.). The right side shows the image’s projection on the learned concept subspace.

Methods	SRCC	PLCC
NIMA [ 85 ]	0.875	0.862
BIAA [ 102 ]	0.858	0.886
HGCN [ 80 ]	0.865	0.881
DINOv2-small [ 66 ]	0.855	0.904
Charm [ 3 ]	0.905	0.938
CLIP-ResNet50 [ 71 ]	0.872	0.903

Table 2: Performance comparison on PARA.

Figure 4: Learned weights of aesthetic concepts on the PARA dataset. The learned bias term is

3.017

Figure 5: Aesthetic score prediction on the PARA test image. Bottom left shows the ground truth and our predictions. The right side shows the image’s projection on the learned concept subspace.

Methods	SRCC	PLCC
LAPIS [ 57 ]	0.809	0.813
Ours (interpretable)	0.805	0.800
Ours (hybrid)	0.817	0.815

Table 3: Performance comparison on LAPIS.

Figure 6: The weights of aesthetic concepts on the LAPIS dataset, we list the four with the highest and lowest weights. For easier understanding, a representative image is included for each one.

Methods	SRCC	PLCC
NIMA [ 85 ]	0.393	0.382
MPADA [ 81 ]	0.437	0.425
MLSP [ 21 ]	0.441	0.430
TANet [ 18 ]	0.453	0.437
SAAN [ 98 ]	0.473	0.467
DINOv2-small [ 66 ]	0.342	0.428

Table 4: Performance comparison on BAID.

Figure 7: Aesthetic score predictions for BAID test images. The bottom shows the four highest concept projection values and their corresponding concept names for each image.

Figure 8: Attribute importance ranking on the AADB dataset obtained by Explainable IAA [ 84 ] .

Explore similar work

AI Outperforms Humans in Personalized Image Aesthetics Assessment via LLM-Based Interviews and Semantic Feature Extraction
May 14, 2026Yoshia Abe, Tatsuya Daikoku, Yasuo KuniyoshiPersonalized Image Aesthetics Assessment Image Quality Assessment
Graduate School of Information Science and Technology The University of Tokyo Bunkyo-ku, Tokyo 113-8654, Japan
One Model, Two Minds: Task-Conditioned Reasoning for Unified Image Quality and Aesthetic Assessment
Mar 20, 2026Wen Yin, Cencen Liu, Dingrui Liu +3Image Quality Assessment Unified Multimodal Models
University of Electronic Science and Technology of China, Chengdu, Sichuan, China · Jiigan Technology, Chengdu, Sichuan, China · Faculty of Information Technology, Monash University, Melbourne, Australia +1
Beyond Absolute Scores: Relative Edit-induced Difference for Generalizable Image Aesthetic Assessment
Jun 4, 2026Qifei Jia, Xintong Yao, Yasen Zhang +7Personalized Image Aesthetics Assessment Image Quality Assessment
Xiaomi Corporation, Beijing, China

May 14, 2026J/K move · Enter open · S save

AI Outperforms Humans in Personalized Image Aesthetics Assessment via LLM-Based Interviews and Semantic Feature Extraction

Yoshia Abe, Tatsuya Daikoku, Yasuo Kuniyoshi

Graduate School of Information Science and Technology The University of Tokyo Bunkyo-ku, Tokyo 113-8654, Japan

Personalized Image Aesthetics Assessment Image Quality Assessment Preference Learning Preference Elicitation

Open paper AskPDF arXiv

Accurately predicting individual aesthetic evaluation for images is a fundamental challenge for AI. Various deep learning (DL)-based models have been proposed for this task, training on image evaluation data to extract objective low-level features. However, aesthetic preferences are inherently subjective and individual-dependent. Accurate prediction thus requires the extraction of high-level semantic features of images and the active collection of preference information from the target individual. To address this issue, we focus on the utility of Large Language Models (LLMs) pretrained on vast amounts of textual data, and develop an integrated DL-LLM system. The system actively elicits aesthetic preferences through LLM-based semi-structured interviews and predicts aesthetic evaluation by leveraging both low-level and high-level features. In our experiments, we compare the proposed system against conventional systems, human predictors, and the target individual's own re-evaluations after a certain time interval. Our results show that the proposed system outperforms all of them, with particularly strong performance on highly-rated images. Moreover, the prediction error of the proposed system is smaller than within-person variability, while human predictors show the largest error, likely due to the influence of their own aesthetic values. These results suggest that AI may be better positioned than others or one's future self to capture individual aesthetic preferences at a given point. This opens a new question of whether AI could serve as a deeper interpreter of human aesthetic sensibility than humans themselves.

Abstract

Explore similar work

AI Outperforms Humans in Personalized Image Aesthetics Assessment via LLM-Based Interviews and Semantic Feature Extraction

One Model, Two Minds: Task-Conditioned Reasoning for Unified Image Quality and Aesthetic Assessment

Beyond Absolute Scores: Relative Edit-induced Difference for Generalizable Image Aesthetic Assessment