PaperlayerView exact revision on arXiv

Source-linked AI summary

ArtiMuse: Fine-Grained Image Aesthetics Assessment with Joint Scoring and Expert-Level Understanding

Shuo Cao, Nan Ma, Jiayang Li, Xiaohui Li, Lihao Shao, Kaiwen Zhu, Yu Zhou, Yuandong Pu, Jiarui Wu, Jiaquan Wang, Bo Qu, Wenhai Wang, Yu Qiao, Dajuin Yao, Yihao Liu

arXiv:2507.14533v2cs.CV

TL;DR

Existing IAA methods and datasets often lack fine-grained, professionally grounded aesthetic analysis alongside scoring. This paper presents ArtiMuse and ArtiMuse-10K, including a token-based approach to continuous scoring, and reports state-of-the-art performance across public aesthetics benchmarks. The authors identify the model’s inability to provide professional aesthetic enhancement recommendations as a limitation.

  • Problem

    Existing IAA methods and datasets often lack fine-grained, professionally curated aesthetic analysis, and MLLMs can struggle to produce objective, expert-aligned judgments.

    Source: §1 · p. 2, §2.1 · p. 3

  • Method

    The paper introduces ArtiMuse for joint aesthetic scoring and expert-level textual analysis, trains it using ArtiMuse-10K, and proposes Token As Score for continuous values.

    Source: §1 · p. 3, §6 · p. 10

  • Results

    ArtiMuse achieves state-of-the-art performance across multiple public aesthetics benchmarks in quantitative assessment and qualitative explanation.

    Source: §1 · p. 3

  • Takeaways & Limitations

    The authors position ArtiMuse as advancing aesthetic AI from score prediction toward holistic reasoning and user-interpretable analysis.

    Source: §1 · p. 3

    The current model can understand and analyze images but cannot provide professional aesthetic enhancement recommendations.

    Source: §6 · p. 10

Abstract

from arXiv · show

The rapid advancement of educational applications, artistic creation, and AI-generated content (AIGC) technologies has substantially increased practical requirements for comprehensive Image Aesthetics Assessment (IAA), particularly demanding methods capable of delivering both quantitative scoring and professional understanding. Multimodal Large Language Model (MLLM)-based IAA methods demonstrate stronger perceptual and generalization capabilities compared to traditional approaches, yet they suffer from modality bias (score-only or text-only) and lack fine-grained attribute decomposition, thereby failing to support further aesthetic assessment. In this paper, we present:(1) ArtiMuse, an innovative MLLM-based IAA model with Joint Scoring and Expert-Level Understanding capabilities; (2) ArtiMuse-10K, the first expert-curated image aesthetic dataset comprising 10,000 images spanning 5 main categories and 15 subcategories, each annotated by professional experts with 8-dimensional attributes analysis and a holistic score. Both the model and dataset will be made public to advance the field.

1 Introduction

ArtiMuse jointly provides fine-grained expert-level textual analysis and aesthetic scores, while ArtiMuse-10K supplies expert annotations across multiple aesthetic attributes. The paper also introduces Token As Score for continuous scoring and reports state-of-the-art performance across public aesthetics benchmarks.

1 Introduction

  • Existing IAA approaches and datasets often lack multidimensional aesthetic analysis and professional, fine-grained annotations, limiting MLLMs’ ability to understand and reason about aesthetics.The introduction distinguishes aesthetic qualities from IQA methods’ focus on technical fidelity and low-level degradations.Source: §1 · p. 2
  • ArtiMuse achieves state-of-the-art performance across multiple public aesthetics benchmarks in quantitative assessment and qualitative explanation.The authors report robust generalization across these benchmarks.Source: §1 · p. 3
  • Token As Score maps existing tokenizer tokens to continuous numeric values, avoiding quantization loss and tokenizer changes.The paper presents it as a lightweight approach for precise, stable score prediction.Source: §1 · p. 3
  • ArtiMuse-10K contains 10,000 images across five main categories and 15 subcategories, with eight expert-described aesthetic attributes and an overall score.The dataset includes textual evaluations and quantitative scores across diverse visual domains.Source: §1 · p. 3
  • ArtiMuse is an MLLM designed to generate fine-grained expert-level textual analysis alongside aesthetic scores.The paper contrasts its joint capabilities with IAA models and general-purpose MLLMs.Source: §1 · p. 3

2 Related Work

Prior work on MLLMs for aesthetics faces challenges in objective, expert-aligned evaluation, while existing datasets and models vary in score coverage, attribute detail, and domain diversity. ArtiMuse is presented as a unified approach to expert-level analysis and aesthetic scoring.

2.1 Multi-modality Large Language Models

  • Current MLLMs can produce overly positive, superficial judgments that differ from professional descriptions, motivating systematic fine-tuning for aesthetic evaluation.The paper notes that their abilities have expanded from image-text matching to high-level semantic understanding.Source: §2.1 · p. 3

2.2 Image Aesthetics Assessment

  • Existing IAA datasets variously lack detailed descriptions, numerical ratings, fine-grained attributes, or coverage beyond photographic content.The paper identifies limited artwork, AIGC, and everyday-scene coverage as additional gaps.Source: §2.2 · p. 3
  • Regression models lack interpretability and generalization, while MLLM-based approaches often lack either granular explanations or score prediction.Q-Align and UNIAA [25] combine text and discrete scores but lack dimension-level evaluation; AesExpert [7] provides expert-style descriptions without scores.Source: §2.2 · p. 4

3 ArtiMuse-10K Dataset

ArtiMuse-10K is an expert-annotated image-aesthetics dataset with broad category coverage and eight fine-grained attributes plus an overall score. It combines curated images, a content-agnostic assessment system, and professional textual annotations.

3.1 Dataset Overview

  • ArtiMuse-10K contains 10,000 images across five main categories and 15 subcategories, each annotated by experts on eight aesthetic attributes and an overall score.The paper describes its diversity and granularity as exceeding existing IAA datasets.Source: §3.1 · p. 4, Table 1 · p. 4

3.2 Image Collection

  • The collection spans Graphic Design, 3D Design, AIGC-generated images, Photography, and Painting & Calligraphy, subdivided into 15 categories including Chinese Painting, Sculpture, and Daily Photography.Non-AIGC images include expert-curated artworks and works from online art and photography platforms; AIGC images include generated and community-contributed samples.Source: §3.2 · p. 4

3.3 Aesthetic Attributes

  • The assessment system defines eight aesthetic attributes and an overall score, and is described as applicable to image types from natural images to AIGC.Attributes include Composition & Design, Technical Execution, Originality & Creativity, Emotion & Viewer Response, and Overall Gestalt.Source: §3.3 · p. 5

3.4 Human Annotations

  • Professional experts annotate each image with textual analysis across eight attributes and an overall aesthetics score.The annotators’ experience ranges from at least three years to over three decades.Source: §3.4 · p. 5

4 Methodology

The methodology builds a diverse image collection, generates annotations according to available source information, and trains ArtiMuse for both aesthetic analysis and scoring. Its scoring strategy maps tokens to score values, while LoRA fine-tuning is used to retain textual capability.

4.1 Dataset Collection & Processing

  • The collection prioritizes existing aesthetic captions and includes images of varying quality, including lower-quality samples, to address diversity and LLM preference bias.It also incorporates more than 350,000 annotated images from existing datasets, including APDDv2 [9], PARA [17], and Impressions [20].Source: §4.1 · p. 5

4.2 Annotation Generation

  • Annotation generation uses score captions for holistic analyses, partial aesthetic captions for eight-attribute evaluations, and expert annotation for professionally selected images.The authors also incorporate human evaluations to counter MLLM positivity bias and provide balanced ground-truth annotations.Source: §4.2 · p. 5

4.3 Training Strategy

  • Training has text-pretraining and score-fine-tuning phases, with LoRA-based LLM fine-tuning used to retain structural analysis ability alongside scoring.The score phase converts each overall aesthetics score into a specialized scoring token; the paper identifies monotonous score data as a risk to textual ability under full LLM fine-tuning.Source: §4.3 · pp. 6–7
  • Token As Score uses existing tokens to represent numeric scores, avoiding vocabulary expansion or tokenizer retraining; the cited comparison reports that 100 aesthetics score tokens achieve optimal results.The paper contrasts this with text-generated scores, which it says can produce severe hallucinations [33] [25].Source: §4.3 · p. 6, Figure 6 · p. 7

5 Experiments

Experiments evaluate ArtiMuse’s aesthetic analysis, scoring, generalization, and design choices. Results report strong benchmark scoring and analysis, with ablations supporting the dataset and training strategy.

5.1 Implementation Details

  • Text pretraining uses batch size 128, learning rate 4e −5, and one epoch; score fine-tuning uses batch size 128, learning rate 2e −5, and two epochs.Experiments use four NVIDIA A100 80GB GPUs; text pretraining takes about five hours, while score fine-tuning takes 10 minutes to four hours depending on dataset size.Source: §5.1 · p. 7

5.2 Structural Aesthetic Analysis

  • ArtiMuse receives a significantly higher human preference rate than other methods, and its analyses outperform them across eight aesthetic attributes.Gemini-2.0-flash judges the eight attributes; volunteers provide the Human Rate.Source: §5.2 · pp. 7–8, Table 2 · p. 8
  • Figure 7 shows ArtiMuse identifying flaws in low-aesthetic images and providing professional assessments of high-aesthetic images, capabilities absent in other models.The qualitative comparison also reports strengths on natural and AIGC images, including compositional cohesion and characteristic AIGC artifacts.Source: Figure 7 · p. 8, §5.2 · p. 8

5.3 Aesthetics Scoring

  • ArtiMuse exceeds other models by over 0.05 PLCC on PARA and ArtiMuse-10K and achieves nearly the highest metrics across scoring datasets.The comparison includes models tested directly and results reported in their original papers.Source: §5.3 · pp. 8–9, Table 3 · p. 9
  • When trained only on AVA, ArtiMuse outperforms Q-Align across all benchmark datasets without adaptation and exceeds several specialized IAA models.Both models are evaluated on out-of-distribution datasets after fine-tuning on AVA alone.Source: §5.3 · p. 9, Table 3 · p. 9

5.4 Ablation Studies

  • Removing any dataset component reduces performance, with the largest drop when Images w/ score caption is excluded; full fine-tuning also harms performance.The authors attribute the full fine-tuning result primarily to loss of aesthetic priors acquired during text pretraining.Source: §5.4 · p. 9, Table 4 · p. 10
  • The two-stage training paradigm outperforms end-to-end joint training, while score-token configuration (i) balances precision and learnability.Insufficient token granularity hampers (g), and excessive token complexity hampers (h).Source: §5.4 · p. 9, Table 4 · p. 10

6 Conclusion

The paper presents ArtiMuse-10K, ArtiMuse, and Token As Score as contributions to image aesthetics assessment. The current model analyzes images but does not provide professional enhancement recommendations.

6 Conclusion

  • ArtiMuse-10K is an expert-annotated image aesthetics dataset, and ArtiMuse provides expert-level textual evaluation and precise aesthetic scoring.Token As Score is described as a lightweight method for precise continuous score prediction in MLLMs.Source: §6 · p. 10
  • The current model is limited to understanding and analyzing images; professional aesthetic enhancement recommendations are left to future work.Source: §6 · p. 10

A ArtiMuse-10K Dataset Details

ArtiMuse-10K uses expert-defined structural annotations across diverse image types. Its appendix describes annotation language, score distributions, category coverage, and public-dataset collection.

A.1 Details of Aesthetic Attributes

  • A domain-expert panel defines eight fine-grained aesthetic attributes, and each image is structurally analyzed across them.Panel members have at least three years of formal aesthetics training, with the most senior having over 30 years of professional experience.Source: §A.1 · p. 15, Table 5 · p. 15

A.2 Charateristics of ArtiMuse-10K

  • The dataset’s frequent annotation terms include “composition” and “visual,” which the authors associate with fundamental visual characteristics in expert assessment.The wordcloud analysis reports that the most frequent terms are strongly correlated with image aesthetic quality.Source: §A.2 · p. 15, Figure 8 · p. 16
  • ArtiMuse-10K contains 9,000 training images and 1,000 test images; the authors report approximately Gaussian score distributions and greater score diversity than other datasets.The score distributions of AVA, PARA, TAD66K, and FLICKR-AES are normalized to [0, 100] for comparison.Source: §A.2 · pp. 15–16
  • Images span five main categories and 15 subcategories, with category counts presented in Table 6.Source: §A.2 · p. 16

B Details of Public Dataset Collection & Processing

B.1 Datasets w/ Score Caption

For public datasets that provide scores without multidimensional annotations, scores guide the MLLM to generate image evaluations. The appendix supplies a prompt and score-distribution figures.

B.1 Datasets w/ Score Caption

  • For datasets with aesthetic scores but no multidimensional annotations, scores guide the MLLM to generate comprehensive image evaluations.Source: §B.1 · p. 16
  • The prompt requests step-by-step analysis of strengths and weaknesses across eight attributes, from composition and technical execution to emotion and comprehensive evaluation.Source: §B.1 · p. 16
  • Figures 9–13 present training and test score distributions for ArtiMuse-10K, AVA, PARA, TAD66K, and FLICKR-AES.Source: Figure 9 · p. 17, Figure 10 · p. 17, Figure 11 · p. 18

B.2 Datasets w/ Partial Text Caption

The section describes how partial textual annotations and aesthetic ratings from existing datasets are filtered and used to create attribute-specific training examples. It then compares score-mapping strategies, finding limitations in five-level discretization and expanded score-token vocabularies.

B.2 Datasets w/ Partial Text Caption

  • APDDv2, Impressions, SPAQ, and KonIQ samples are filtered or selected and paired with attribute-specific prompts that combine available captions, comments, categories, or scores.The selected sources include 4,898 APDDv2 instances, 1,443 Impressions annotations, and 1,488 KonIQ images; SPAQ selection emphasizes high- and low-scoring examples.Source: §B.2 · pp. 20–21

C Details of Token As Score Strategy

  • Across AVA and ArtiMuse-10K, experiments vary score-prediction methodology while holding training data, configurations, and model architecture consistent.The comparison is designed to isolate the score-prediction strategy.Source: C Details of Token As Score Strategy · p. 22

C.1 Level As Score

  • Five-level Level As Score underperforms Token As Score, with coarse partitions and vocabulary misalignment identified as contributors to the degradation.Level As Score maps continuous scores to five discrete levels and reconstructs predictions using their probability distribution.Source: §C.1 · p. 22

C.2 Token As Score w/ Expanding Tokens

  • Expanding score-token vocabularies peaks at 100 tokens on AVA, but performs poorly on ArtiMuse-10K, where the approach may fail to converge.The authors attribute the AVA trend to insufficient mapping granularity at low token counts and underfitting when token complexity exceeds available data or model capacity.Source: §C.2 · pp. 22–23

C.3 Token As Score w/ Existing Tokens

This section compares existing displayable tokens for score prediction and adopts an ordered 100-token mapping in ArtiMuse. Existing tokens outperform newly introduced tokens, and performance improves as the number of existing tokens increases.

C.3 Token As Score w/ Existing Tokens

  • ArtiMuse uses 100 existing tokens formed from lowercase letters and ordered in the Qwen2.5-7B vocabulary, mapped sequentially to scores from 0 to 100.The compared configurations include 25-token and 50-token mappings, as well as ordered and non-ordered 100-token mappings.Source: §C.3 · p. 23, Table 9 · p. 23
  • Existing-vocabulary tokens consistently outperform newly introduced tokens at matched token counts, while performance improves as existing-token counts increase.The authors attribute the advantage to pretrained knowledge and clearer semantic and ordinal relationships; the ordered mapping in (j) also performs better than (i).Source: §C.3 · pp. 23–24, Table 10 · p. 24

D.1 Training Details

The supplementary sections detail ArtiMuse’s training, evaluation procedures, benchmark comparisons, and image examples. Reported results include stronger aesthetics scoring than leading MLLMs and better generalization than Q-Align on unseen datasets in most cases.

D.1 Training Details

  • Fixed-resolution processing improves SRCC and PLCC by approximately 0.3 each over dynamic high-resolution processing and more than doubles training and inference efficiency.The authors motivate avoiding image tiling because it can fragment global composition, color harmony, and spatial relationships.Source: §D.1 · p. 25

D.2 Inference Details for Aesthetics Scoring

  • For scoring-capable models, evaluation uses their score outputs; models without scoring ability receive prompts to return an integer from 0 to 100.Prompted scores are used for comparative analysis in ArtiMuse’s scoring format.Source: §D.2 · p. 25

D.3 Inference Details for Textual Analysis

  • Text-analysis comparisons prompt models to assess a specified aesthetic attribute, providing other models with the attribute description for contextual parity.The prompts limit responses to 100 words.Source: §D.3 · p. 25

D.4 Comparison Details

  • A judgment MLLM selects the response best aligned with human expert critique, while a separate volunteer study collects model preferences across 20 images and eight attributes.The judgment prompt assesses clarity, accuracy, insightfulness, and relevance; the user study recruits 20 volunteers with varied aesthetic-evaluation experience.Source: §D.4 · pp. 26–27, Figure 14 · p. 26

E.1 Comparison with SOTA Open-Source & Closed-Source MLLMs

  • ArtiMuse achieves higher aesthetics-scoring performance than leading open- and closed-source MLLMs despite having only 8B parameters.The comparison includes open-source Qwen-2.5-VL-72B-instruct and InternVL3-78B, and closed-source GPT-4o and Gemini-2.0-Flash.Source: §E.1 · p. 27, Table 12 · p. 27

E.2 Further Comparison of Generalization Ability

  • ArtiMuse trained on five datasets outperforms Q-Align on unseen datasets in most cross-dataset evaluations.The evaluation trains both models on AVA, PARA, TAD66K, FLICKR-AES, and ArtiMuse-10K, then tests across all five datasets.Source: §E.2 · p. 27, Table 13 · p. 27

E.3 Image Examples in ArtiMuse-10K; E.4 Complete Examples in ArtiMuse-10K; E.5 Further Comparison of Textual Analysis

  • The supplementary examples span ArtiMuse-10K’s categories, show expert assessments across eight attributes, and include textual analyses on test-set and real-world images.The examples cover Photography, Painting & Calligraphy, AIGC, 3D Design, and Graphic Design; real-world examples compare model analyses with expert commentary.Source: §E.3 · p. 27, §E.4 · p. 28, §E.5 · p. 28

E.6 Results on Real-world Images

ArtiMuse is evaluated on real-world images as an out-of-distribution setting. The authors report that it maintains accurate, expert-level analyses that identify visual strengths and weaknesses.

E.6 Results on Real-world Images

  • On real-world out-of-distribution images, ArtiMuse maintains accurate, expert-level aesthetic analyses that identify strengths and weaknesses.The examples are shown in Figures 29–31.Source: §E.6 · p. 28

Children's Painting

This section presents expert-style assessments of children’s paintings, pairing eight-dimensional attribute analysis with holistic scores. Across examples, it characterizes strengths and limitations in composition, technique, originality, communication, creativity, and viewer response, including attribute-specific and real-world analysis.

Children's Painting

  • 52: The plum-blossom photograph has appealing colors and a clear spring theme, but its blurred buildings disrupt harmony and its concept is conventional.The focused buds, exposure, and depth-of-field effect are handled well, though the background is fuzzy and the image offers limited emotional depth.Source: Children's Painting · p. 32
  • 95: The rabbit painting is judged an outstanding, highly complete naturalistic work, with exceptional technique, vivid detail, and strong emotional resonance.Its evaluation also emphasizes a balanced composition, originality, and harmony, describing it as a model of realistic painting with important historical and artistic value.Source: Children's Painting · p. 33
  • 21: The horse painting is assessed as a basic modeling exercise, with immature technique, weak structure and volume, and little emotional resonance.Its recognizable subject and regular composition do not compensate for limited originality, bland artistry, and insufficiently developed details.Source: Children's Painting · p. 34
  • 62: The flamingo-and-child image is praised for its imaginative surreal concept, but AI traces and distortions of reality limit its artistic evaluation.Its soft contrasting colors and distinctive scene evoke curiosity, while some element combinations and character details remain visibly imperfect.Source: Children's Painting · p. 35
  • 92: The product image earns a high score for professional execution, harmonious color, and a clear, commercially appealing presentation.The evaluation highlights its composition and technical quality, while noting that its originality is expressed mainly through the product design.Source: Children's Painting · p. 36
  • The tire design is presented as a technical sketch whose basic rendering, jagged edges, and background grid weaken its overall impression.Its distinctive tread offers some originality, but the conventional presentation, limited emotional engagement, and low artistry constrain the work.Source: Children's Painting · p. 37
  • A separate technical image received an overall aesthetics score of 28; its design concept was evident, but rendering, color, depth, and emotional arousal were limited.The assessment judged its artistry in need of improvement.Source: Children's Painting · p. 37
  • A donation-themed image received an overall aesthetics score of 52, with clear communication and creative appeal offset by average completeness and static composition.Its mixed-media contrast concretizes donation as building a home, supports emotional resonance, and serves its publicity purpose, though technical fusion remains basic.Source: Children's Painting · p. 38

Photography Comparison on Attribute 5. Theme & Communication

  • On photographic images from the ArtiMuse-10K test set, ArtiMuse showed superior performance identifying flaws in low-aesthetic examples, with higher accuracy than competing models and close alignment with expert human evaluations.Source: Figure 27 · p. 40

Painting & Calligraphy Comparison on Attribute 4. Originality & Creativity

  • On paintings in the ArtiMuse-10K test set, ArtiMuse showed superior technical analysis of high-aesthetic images, with higher accuracy than competing models and near-expert-level performance in aesthetic evaluation.Source: Figure 28 · p. 41
  • Real-world image examples show ArtiMuse producing granular assessments of visual content and evaluations of both strengths and weaknesses.The examples are described as expert-level analyses.Source: Figure 29 · p. 42, Figure 30 · p. 42, Figure 31 · p. 43

Related papers

Loading 2507.14533v2…