Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

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

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read ArtiMuse pairs expert-level aesthetic critique with high-accuracy scoring in a single multimodal model.

desk verdict The dataset is the real contribution; the SOTA scoring claim is overstated, and the 'expert-level understanding' evaluation is the soft underbelly. read the letter →

arxiv 2507.14533 v2 pith:VGXYTU55 submitted 2025-07-19 cs.CV

Shuo Cao , Nan Ma , Jiayang Li , Xiaohui Li , Lihao Shao , Kaiwen Zhu , Yu Zhou , Yuandong Pu
show 7 more authors
Jiarui Wu Jiaquan Wang Bo Qu Wenhai Wang Yu Qiao Dajuin Yao Yihao Liu
This is my paper · ORCID
classification cs.CV
keywords ImageAestheticsAssessmentMultimodalLargeLanguageModelFine-grainedaestheticattributesExpertannotationdatasetScorepredictionTokenAsAIGCqualitycritiquegeneration
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that image aesthetics assessment can be done well by a single multimodal model that both writes expert-level, attribute-by-attribute critiques and predicts a holistic score. To support this, the authors built ArtiMuse-10K, a 10,000-image dataset annotated by professional artists and art scholars along eight defined aesthetic attributes plus an overall score, and trained ArtiMuse, an 8B-parameter MLLM, on it. They report state-of-the-art or near-best Spearman and Pearson correlations on AVA, PARA, TAD66K, FLICKR-AES, and their own test split, and a 71.1% win rate when a judge MLLM compares its critiques against those of other models using an expert critique as reference. The paper's contribution is thus both a new benchmark-style dataset and a method for continuous score prediction inside an MLLM without adding vocabulary tokens.

What carries the argument

The load-bearing mechanism is Token As Score: using existing twin-letter tokens of the LLM tokenizer, in their natural order, as dense numeric encodings of scores 0–100, then predicting a token distribution and taking its expectation. This carries the scoring claim by avoiding both free-form text regression (which the paper says hallucinates) and coarse five-level discretization (which loses precision). The second machinery is the fixed eight-attribute annotation schema — Composition & Design, Visual Elements & Structure, Technical Execution, Originality & Creativity, Theme & Communication, Emotion & Viewer Response, Overall Gestalt, Comprehensive Evaluation — that structures both the expert annotations and the model's textual output, making the critique fine-grained and comparable.

What would settle it

Re-annotate a random 200-image subset of ArtiMuse-10K with independent expert annotators using the same eight-attribute rubric, and compute inter-annotator agreement on both critiques and holistic scores; if agreement is low (e.g., ICC below 0.5 on scores or low semantic similarity on critiques), then the dataset's labels and the evaluation built on them do not fix a reliable 'expert' standard. Alternatively, replace the Gemini-2.0-flash judge with a panel of professional art critics and see whether ArtiMuse still wins by 71%.

Watch

Extended reading notes

Core claim

The central claim, stated on the paper's own terms, is that ArtiMuse achieves expert-level aesthetic understanding and precise aesthetics scoring at the same time, and that this joint ability transfers across photographic, artistic, design, and AI-generated content. On the scoring side, the paper's Token As Score strategy maps the integers 0–100 onto 101 existing, ordered character tokens in the LLM's vocabulary (e.g., 'aa','ab',...), trains the model to emit the token for the correct score, and at inference computes the final score as the expectation of the predicted token distribution. On the understanding side, the model is trained to generate structured, eight-attribute textual evaluations that are judged to be substantially better than those of AesExpert, Qwen-2.5-VL, and InternVL-3, both by a judge MLLM and by volunteers.

Load-bearing premise

The claim of 'expert-level understanding' rests on the assumptions that the one expert critique used per image is ground truth and that Gemini-2.0-flash's choice of best model response, plus a 20-image human vote, accurately measures analysis quality.

Editorial extensions

If this is right

  • A single 8B MLLM can carry both dense regression-style scoring and multi-attribute textual reasoning without catastrophic forgetting, if trained with the paper's two-stage LoRA schedule.
  • Aesthetics benchmarks can be compared head-to-head even when label distributions differ, because Token As Score predictions are on a 0–100 scale and match or beat dedicated regressors.
  • The ArtiMuse-10K dataset gives the community 10,000 expert-written, attribute-structured critiques with scores, enabling training of models that can say why an image is weak or strong, not just how weak or strong.
  • Models trained with this recipe should generalize zero-shot to unseen aesthetics datasets better than existing MLLM-based scorers, as shown by the AVA-trained transfer experiments.
  • Since only existing tokens are used, the strategy is drop-in compatible with any tokenizer and does not require vocabulary expansion or embedding resizing.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the 71.1% judge-MLLM win rate reflects genuine expert-level critique quality, then ArtiMuse-style models could be used as low-cost first-pass critics in photography education, design review, or AIGC curation, flagging specific compositional and technical flaws before a human expert looks at the image.
  • The Token As Score idea is not limited to aesthetics: any continuous regression target inside an LLM (e.g., quality scores, depth estimates, moral judgments on a scale) could be mapped onto ordered existing tokens, provided the tokenizer has enough low-entropy ordered tokens; the paper's ablation suggests token count and ordinality matter more than the specific vocabulary.
  • The reliance on a single expert critique per image as ground truth is the point that would benefit most from replication: an independent re-annotation study, even on a few hundred images, would tell whether the eight-attribute schema supports reproducible 'expert-level' judgments or mostly measures one expert's style.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper introduces ArtiMuse, an InternVL-3-8B based multimodal large language model for joint image aesthetics scoring and fine-grained textual analysis across eight aesthetic attributes, together with ArtiMuse-10K, a 10,000-image expert-annotated dataset spanning five main categories, and a 'Token As Score' strategy that maps continuous aesthetic scores to existing LLM tokens. The model is trained in two stages on a mixture of public datasets and ArtiMuse-10K, and it is evaluated on AVA, PARA, TAD66K, FLICKR-AES, and ArtiMuse-10K for scoring, and on an MLLM-judge and a small human preference study for textual analysis. The central claims are that ArtiMuse simultaneously achieves state-of-the-art scoring performance and expert-level structural aesthetic understanding.

Significance. If the claims are supported, the paper would provide a useful new resource and model: ArtiMuse-10K is the first expert-annotated dataset with eight fine-grained aesthetic attributes and holistic scores, the Token As Score strategy is a lightweight and reproducible way to improve continuous score prediction with MLLMs, and the model shows strong results on several benchmarks, particularly FLICKR-AES and PARA. The paper also reports ablations on dataset composition, training strategy, and score-token granularity, which are informative. However, the claimed 'significantly outperforming' and 'expert-level' results are not fully established by the presented evidence.

major comments (4)
  1. [§5.3, Table 3] The headline claim that ArtiMuse 'significantly outperforms' state-of-the-art models is not supported by the reported margins. On AVA, ArtiMuse's SRCC is 0.827, which is lower than Next Token Is Enough's 0.828, and on TAD66K its SRCC is 0.510, lower than AesMamba's 0.511; the PLCC differences on these datasets are at most 0.012. No statistical significance tests, confidence intervals, or multiple-seed variances are reported, and results for some baselines are taken directly from their original papers while ArtiMuse is fine-tuned per target dataset. The authors should either temper the 'significantly outperforming' language to 'competitive' or provide statistical evidence and a fully matched evaluation protocol.
  2. [§5.2 and Appendix D.4] The expert-level textual analysis claim rests on an evaluation yardstick that is not validated. Each ArtiMuse-10K image has exactly one expert-written critique, and no inter-annotator agreement is reported for these critiques. The MLLM-judge protocol asks Gemini-2.0-flash to select the response best aligned with that single human critique, but no evidence is given that Gemini's preferences track independent human expert judgments. The human confirmation study uses only 20 images and 20 volunteers, with no confidence intervals or expert/non-expert stratification. Without inter-annotator agreement, judge-model validation, or a larger human study, the reported 71.1% MLLM-judge rate and 67.8% human preference rate do not establish expert-level understanding.
  3. [§5.4, Table 4 and Appendix C] The ablation study selects the 100-token ordered mapping and other hyperparameters using the same benchmarks (AVA and ArtiMuse-10K) that are then used in the main state-of-the-art Table 3, introducing a benchmark-selection effect. For example, the gains of the ordered 100-token strategy over the non-ordered variant are 0.001 SRCC on AVA but 0.032 SRCC on ArtiMuse-10K, and the choice of the ordered mapping is inferred from these two datasets. The authors should validate the final configuration on a held-out benchmark not used during ablation, or at least explicitly acknowledge that the reported gains may be partly due to selection on the evaluation sets.
  4. [§5.3, Table 3] The comparison with baselines is not fully controlled. For TANet, AesMamba, UNIAA-LLaVA, and Next Token Is Enough, results are taken from original papers, while Q-Align and ArtiMuse are fine-tuned on each target dataset; for models without scoring capability, scores are elicited by prompting. The exact training splits, score normalization, and prompt templates can differ across methods, and the paper does not report whether all methods use the same test splits. A fairer comparison, or at least a detailed description of the shared evaluation protocol for every baseline, is needed before claiming superiority.
minor comments (5)
  1. [Abstract and §5.3] The abstract and introduction state that ArtiMuse 'significantly outperforms' state-of-the-art models, but Table 3 shows that ArtiMuse is behind on AVA SRCC and TAD66K SRCC; the wording should be aligned with the actual numerical results.
  2. [Table 1] Several entries in Table 1 contain malformed symbols such as '/times' and '/times–', which should be replaced by proper check and cross marks for readability.
  3. [Appendix C.3] The text refers to the 'Qwen2.5-7B LLM' vocabulary, while the base model is InternVL-3-8B with a Qwen2.5-7B language backbone; this relationship should be stated clearly to avoid confusion about which tokenizer is used.
  4. [§6 Limitations] The Limitations section only mentions the absence of enhancement recommendations; it should also acknowledge the single-expert annotation protocol and the lack of validation of the MLLM judge, which are more immediate threats to the expert-level claim.
  5. [Throughout] There are typographical errors such as 'ArtMuse' in the contributions list, 'AriMuse' in Appendix C, 'futher' in Appendix C.3, and 'T otal' in Table 6; these should be corrected.

Circularity Check

1 steps flagged · score 4.0 of 10

Score-token configuration is selected on the same AVA and ArtiMuse-10K benchmarks that later appear in the SOTA table, so part of the reported advantage on those two datasets is a selection artifact; other benchmarks give independent support.

  1. fitted input called prediction [Appendix C (Token As Score exploration) and Section 5.4, reported in Section 5.3 Table 3]
    "To ensure robust and reliable experimental conclusions, we conduct comprehensive evaluations on both A V A (the largest image aesthetics scoring dataset) [18] and ArtiMuse-10K (ours)."

    The score-prediction strategy was chosen by comparing configurations on A V A and ArtiMuse-10K; Table 10 selects the 100 ordered tokens as the best on those exact datasets, and Section 5.4 calls configuration (i) 'our final choice.' Table 3 then reports this same configuration as ArtiMuse's SOTA on A V A (SRCC 0.827, PLCC 0.826) and ArtiMuse-10K (SRCC 0.614, PLCC 0.627). The claimed advantage on these two benchmarks is therefore partly a product of hyperparameter selection on the evaluation set rather than an independent out-of-sample prediction. This is a partial benchmark loop rather than a full equation-level reduction, and the gains on PARA, TAD66K, and FLICKR-AES, which were not used for this selection, provide independent evidence for the scoring method.

full rationale

ArtiMuse is an empirical MLLM paper rather than a formal derivation, so the circularity check focuses on benchmark loops and self-citation chains. The Token As Score pipeline is a concrete modeling choice (Section 4.3 and Appendix C): scores are mapped to 100 existing tokens and reconstructed as an expectation over token probabilities; this is tested in ablations and does not reduce by definition to the benchmark labels. The authors' own prior works [29,30,31] are cited only as related-work support for dataset diversity and are not load-bearing. The main defensible circularity flag is that the final score-token configuration was selected using A V A and ArtiMuse-10K, the same two datasets highlighted in the SOTA table; the exact AVA and ArtiMuse-10K numbers are therefore partly a selection artifact. I do not treat the Gemini-2.0-flash judging protocol as circular: evaluating on held-out images against a single expert critique is a generalization test, and its weakness is construct validity (no judge validation, no inter-annotator agreement) rather than derivation-from-input. The Limitations section only notes the missing enhancement-recommendation capability, so the paper does not disclose the benchmark-selection loop, but that omission does not change the severity. Because the central scoring capability retains independent support on three public benchmarks, the overall score is 4.

Assumptions & free parameters 5 free parameters · 5 assumptions · 1 invented entities

The central claims rest on the expert quality of ArtiMuse-10K annotations, the validity of the 8-attribute rubric, and the benchmarks used to measure both scoring and text quality. These are domain assumptions plus one author-defined rubric, not standard math axioms. The main free parameters are the score-token granularity and training choices selected by ablation on the same datasets used for final results.

free parameters (5)
  • Number of aesthetics score tokens = 100
    Selected as optimal in Table 10 by SRCC/PLCC on AVA and ArtiMuse-10K; the central scoring claim depends on this granularity.
  • Score token vocabulary mapping = Existing tokens aa through ey mapped to scores 0 through 100
    Appendix C.3; authors chose twin-letter pairs with ordinal properties, but no evidence is given that this mapping is uniquely good.
  • LoRA rank = 16 for text pretraining, 128 for score finetuning
    Appendix D.1; no sensitivity study is reported, and rank affects the text and score trade-off.
  • Image resolution strategy = Fixed 448x448, no tiling
    Appendix D.1; authors report about 0.3 SRCC/PLCC gain over dynamic resolution, but this is a hand-chosen modification.
  • Learning rates and epochs = 4e-5 for 1 epoch; 2e-5 for 2 epochs
    Appendix D.1; standard choices, not swept or justified by sensitivity analysis.
assumptions (5)
  • domain assumption Expert annotations are treated as reliable ground truth.
    Section 3.4 and Appendix A.1 describe experts with 3 to 30 plus years of experience, but no inter-annotator agreement, annotation protocol details, or repeated annotations are reported.
  • ad hoc to paper The 8-attribute taxonomy is a valid, content-agnostic decomposition of aesthetics.
    Section 3.3 defines the taxonomy through expert consultation; the paper provides no external validation that these attributes, rather than some other set, capture aesthetic quality.
  • domain assumption Gemini-2.0-flash preference, guided by one expert critique, is a valid measure of textual analysis quality.
    Section 5.2 and Appendix D.4 use a single MLLM judge for the main textual claim; no correlation with human expert agreement is reported.
  • domain assumption Cross-dataset score normalization to 0 to 100 preserves ordinal comparability.
    Section 4.3 and Appendix A.2 normalize AVA, PARA, and other scores; if distributions are not comparable, joint training and transfer numbers are distorted.
  • domain assumption Fixed-resolution resizing preserves the global aesthetic features needed for scoring.
    Appendix D.1 replaces dynamic tiling with 448x448 resizing; authors say it improves performance but provide no diagnostic evidence.
invented entities (1)
  • Eight-attribute aesthetic rubric
    purpose: Defines the annotation schema, model output format, and evaluation dimensions for ArtiMuse-10K and ArtiMuse.
    The rubric is introduced by the authors' expert consultations and is not validated against external aesthetic theory or inter-annotator agreement; it is a new conceptual entity that shapes every claim in the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ArtiMuse: Fine-Grained Image Aesthetics Assessment with Joint Scoring and Expert-Level Understanding." pith.science (2026). https://pith.science/paper/VGXYTU55

@misc{pith2026250714533,
  author       = {Pith},
  title        = {Pith review of: ArtiMuse: Fine-Grained Image Aesthetics Assessment with Joint Scoring and Expert-Level Understanding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VGXYTU55}},
  note         = {Machine review of arXiv:2507.14533}
}
read the original abstract

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.

Figures

Figures reproduced from arXiv: 2507.14533 by the authors.

Figure 1
Figure 1. ArtiMuse provides granular, expert-level textual understanding results for images across [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. In comparison with existing models, ArtiMuse outperforms them by simultaneously [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Data examples in ArtiMuse-10K [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (27 more)
Figure 5
Figure 5. Figure 5: Overview of ArtiMuse. ArtiMuse encompasses a multi-stage pipeline spanning data [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Comparison of score prediction methods. Token As Score features a more rational design [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Structural aesthetic analysis results. Red, green, and brown denote positive, negative, and expert-level analyses, respectively. ArtiMuse uniquely identifies flaws in low-aesthetic images while providing professional assessment of high-aesthetic images, capabilities ab…
Figure 8
Figure 8. Figure 8: Wordcloud of ArtiMuse-10K dataset. The most frequent words in ArtiMuse-10K dataset [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]
Figure 9
Figure 9. Figure 9: Score distribution of training and test splits in ArtiMuse-10K. [PITH_FULL_IMAGE:figures/full_fig_p017_9.png]
Figure 10
Figure 10. Figure 10: Score distribution of training and test splits in AVA. [PITH_FULL_IMAGE:figures/full_fig_p017_10.png]
Figure 11
Figure 11. Figure 11: Score distribution of training and test splits in PARA. [PITH_FULL_IMAGE:figures/full_fig_p018_11.png]
Figure 12
Figure 12. Figure 12: Score distribution of training and test splits in TAD66K. [PITH_FULL_IMAGE:figures/full_fig_p018_12.png]
Figure 13
Figure 13. Figure 13: Score distribution of training and test splits in FLICKR-AES. [PITH_FULL_IMAGE:figures/full_fig_p019_13.png]
Figure 14
Figure 14. Figure 14: Pipeline of the structural aesthetic analysis ability judgment by MLLM. [PITH_FULL_IMAGE:figures/full_fig_p026_14.png]
Figure 15
Figure 15. Figure 15: Image examples from the Photography category in ArtiMuse-10K dataset. 29 [PITH_FULL_IMAGE:figures/full_fig_p029_15.png]
Figure 16
Figure 16. Figure 16: Image examples from the Painting & Calligraphy category in ArtiMuse-10K dataset. 30 [PITH_FULL_IMAGE:figures/full_fig_p030_16.png]
Figure 17
Figure 17. Figure 17: Image examples from the AIGC, 3D Design and Graphic Design categories in ArtiMuse￾10K dataset. 1.Composition & Design: The composition is standard and effective. The diagonal line of the bridge is used to guide the line of sight. The bridge in the foreground, the trad…
Figure 18
Figure 18. Figure 18: High-aesthetic example from Photography category. 31 [PITH_FULL_IMAGE:figures/full_fig_p031_18.png]
Figure 19
Figure 19. Figure 19: Medium-aesthetic example and low-aesthetic example from [PITH_FULL_IMAGE:figures/full_fig_p032_19.png]
Figure 20
Figure 20. Figure 20: High-aesthetic and medium-aesthetic example from [PITH_FULL_IMAGE:figures/full_fig_p033_20.png]
Figure 21
Figure 21. Figure 21: Low-aesthetic example from Painting & Calligraphy category and high-aesthetic example from AIGC category. 34 [PITH_FULL_IMAGE:figures/full_fig_p034_21.png]
Figure 22
Figure 22. Figure 22: Medium-aesthetic example and low-aesthetic example from [PITH_FULL_IMAGE:figures/full_fig_p035_22.png]
Figure 23
Figure 23. Figure 23: High-aesthetic and medium-aesthetic example from [PITH_FULL_IMAGE:figures/full_fig_p036_23.png]
Figure 24
Figure 24. Figure 24: Low-aesthetic example from 3D Design category and high-aesthetic example from Graphic Design category. 37 [PITH_FULL_IMAGE:figures/full_fig_p037_24.png]
Figure 25
Figure 25. Figure 25: Medium-aesthetic example and low-aesthetic example from [PITH_FULL_IMAGE:figures/full_fig_p038_25.png]
Figure 26
Figure 26. Figure 26: Structural aesthetic analysis results and comparisons on AIGC images from ArtiMuse-10K [PITH_FULL_IMAGE:figures/full_fig_p039_26.png]
Figure 27
Figure 27. Figure 27: Structural aesthetic analysis results and comparisons on photographic images from [PITH_FULL_IMAGE:figures/full_fig_p040_27.png]
Figure 28
Figure 28. Figure 28: Structural aesthetic analysis results and comparisons on paintings from ArtiMuse-10K test [PITH_FULL_IMAGE:figures/full_fig_p041_28.png]
Figure 29
Figure 29. Figure 29: Textual results of ArtiMuse on real-world images. [PITH_FULL_IMAGE:figures/full_fig_p042_29.png]
Figure 30
Figure 30. Figure 30: Textual results of ArtiMuse on real-world images. [PITH_FULL_IMAGE:figures/full_fig_p042_30.png]
Figure 31
Figure 31. Figure 31: Textual results of ArtiMuse on real-world images. [PITH_FULL_IMAGE:figures/full_fig_p043_31.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. UltraFlux: Data-Model Co-Design for High-quality Native 4K Text-to-Image Generation across Diverse Aspect Ratios

    cs.CV 2025-11 conditional novelty 5.0 of 10

    A data-model co-designed Flux DiT with resonance RoPE, post-trained F16 VAE, SNR-aware Huber wavelet loss, and aesthetic curriculum claims state-of-the-art open-source 4K multi-AR text-to-image generation.

Reference graph

Works this paper leans on

43 extracted references · 29 canonical work pages · cited by 1 Pith paper

  1. [1]

    Dreamlike photoreal 2.0,

    dreamlike.art, “Dreamlike photoreal 2.0,” https://huggingface.co/dreamlike-art/dreamlike-photoreal-2.0, 2023, accessed: 2025-05-14, Licensed under CreativeML OpenRAIL-M (modified)

    work page 2023

  2. [2]

    B. F. Labs, “Flux,” https://github.com/black-forest-labs/flux, 2024

    2024

  3. [3]

    High-resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High-resolution image synthesis with latent diffusion models,” 2021

    2021

  4. [4]

    Depicting beyond scores: Advancing image quality assessment through multi-modal language models,

    Z. You, Z. Li, J. Gu, Z. Yin, T. Xue, and C. Dong, “Depicting beyond scores: Advancing image quality assessment through multi-modal language models,” inEuropean Conference on Computer Vision, 2024

    work page 2024

  5. [5]

    Descriptive image quality assessment in the wild,

    Z. You, J. Gu, Z. Li, X. Cai, K. Zhu, C. Dong, and T. Xue, “Descriptive image quality assessment in the wild,”arXiv preprint arXiv:2405.18842, 2024

    arXiv 2024

  6. [6]

    Teaching large language models to regress accurate image quality scores using score distribution,

    Z. You, X. Cai, J. Gu, T. Xue, and C. Dong, “Teaching large language models to regress accurate image quality scores using score distribution,” inIEEE Conference on Computer Vision and Pattern Recognition, 2025

    work page 2025

  7. [7]

    AesExpert: Towards Multi-modality Foundation Model for Image Aesthetics Perception

    Y . Huang, X. Sheng, Z. Yang, Q. Yuan, Z. Duan, P. Chen, L. Li, W. Lin, and G. Shi, “Aesexpert: Towards multi-modality foundation model for image aesthetics perception,”arXiv:2404.09624, 2024

    work page Pith review arXiv 2024

  8. [8]

    Aesmamba: Universal image aesthetic assessment with state space models,

    F. Gao, Y . Lin, J. Shi, M. Qiao, and N. Wang, “Aesmamba: Universal image aesthetic assessment with state space models,” inProceedings of the 32nd ACM International Conference on Multimedia, 2024, pp. 7444–7453

    work page 2024

Show all 43 references
  1. [9]

    Apddv2: Aesthetics of paintings and drawings dataset with artist labeled scores and comments,

    X. Jin, Q. Qiao, Y . Lu, H. Wang, H. Huang, S. Gao, J. Liu, and R. Li, “Apddv2: Aesthetics of paintings and drawings dataset with artist labeled scores and comments,” 2024. [Online]. Available: https://arxiv.org/abs/2411.08545

    2024 arXiv

  2. [10]

    Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models,

    J. Zhu, W. Wang, Z. Chen, Z. Liu, S. Ye, L. Gu, Y . Duan, H. Tian, W. Su, J. Shaoet al., “Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models,”arXiv preprint arXiv:2504.10479, 2025

    2025 arXiv

  3. [11]

    Qwen2.5-vl technical report,

    S. Bai, K. Chen, X. Liu, J. Wang, W. Ge, S. Song, K. Dang, P. Wang, S. Wang, J. Tanget al., “Qwen2.5-vl technical report,”arXiv preprint arXiv:2502.13923, 2025

    2025 arXiv

  4. [12]

    Q-align: Teaching LMMs for visual scoring via discrete text-defined levels,

    H. Wu, Z. Zhang, W. Zhang, C. Chen, L. Liao, C. Li, Y . Gao, A. Wang, E. Zhang, W. Sun, Q. Yan, X. Min, G. Zhai, and W. Lin, “Q-align: Teaching LMMs for visual scoring via discrete text-defined levels,” inProceedings of the 41st International Conference on Machine Learning, se...

    2024

  5. [13]

    Gpt-4 technical report,

    J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkatet al., “Gpt-4 technical report,”arXiv preprint arXiv:2303.08774, 2023. 10

    2023 arXiv

  6. [14]

    Gemini: a family of highly capable multimodal models,

    G. Team, R. Anil, S. Borgeaud, J.-B. Alayrac, J. Yu, R. Soricut, J. Schalkwyk, A. M. Dai, A. Hauth, K. Mil- licanet al., “Gemini: a family of highly capable multimodal models,”arXiv preprint arXiv:2312.11805, 2023

    2023 arXiv

  7. [15]

    Rethinking image aesthetics assessment: Models, datasets and benchmarks,

    S. He, Y . Zhang, R. Xie, D. Jiang, and A. Ming, “Rethinking image aesthetics assessment: Models, datasets and benchmarks,”IJCAI, 2022

    2022

  8. [16]

    Photo aesthetics ranking network with attributes and content adaptation,

    S. Kong, X. Shen, Z. Lin, R. Mech, and C. Fowlkes, “Photo aesthetics ranking network with attributes and content adaptation,” 2016. [Online]. Available: https://arxiv.org/abs/1606.01621

    2016 arXiv

  9. [17]

    Personalized image aesthetics assessment with rich attributes,

    Y . Yang, L. Xu, L. Li, N. Qie, Y . Li, P. Zhang, and Y . Guo, “Personalized image aesthetics assessment with rich attributes,” 2022. [Online]. Available: https://arxiv.org/abs/2203.16754

    2022 arXiv

  10. [18]

    Ava: A large-scale database for aesthetic visual analysis,

    N. Murray, L. Marchesotti, and F. Perronnin, “Ava: A large-scale database for aesthetic visual analysis,” in 2012 IEEE Conference on Computer Vision and Pattern Recognition, 2012, pp. 2408–2415

    2012

  11. [19]

    Artemis: Affective language for visual art,

    P. Achlioptas, M. Ovsjanikov, K. Haydarov, M. Elhoseiny, and L. Guibas, “Artemis: Affective language for visual art,”CoRR, vol. abs/2101.07396, 2021

    2021 arXiv

  12. [20]

    Impressions: Understanding visual semiotics and aesthetic impact,

    J. Kruk, C. Ziems, and D. Yang, “Impressions: Understanding visual semiotics and aesthetic impact,”

  13. [21]

    Understanding aesthetics with language: A photo critique dataset for aesthetic assessment,

    D. V . Nieto, L. Celona, and C. F. Labrador, “Understanding aesthetics with language: A photo critique dataset for aesthetic assessment,” inThirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. [Online]. Available: https://openre...

    2022

  14. [22]

    Towards artistic image aesthetics assessment: a large-scale dataset and a new method,

    R. Yi, H. Tian, Z. Gu, Y .-K. Lai, and P. L. Rosin, “Towards artistic image aesthetics assessment: a large-scale dataset and a new method,” 2023. [Online]. Available: https://arxiv.org/abs/2303.15166

    2023 arXiv

  15. [23]

    Q- instruct: Improving low-level visual abilities for multi-modality foundation models,

    H. Wu, Z. Zhang, E. Zhang, C. Chen, L. Liao, A. Wang, K. Xu, C. Li, J. Hou, G. Zhaiet al., “Q- instruct: Improving low-level visual abilities for multi-modality foundation models,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 2...

    2024

  16. [24]

    Scaling up personalized image aesthetic assessment via task vector customization,

    J. Yun and J. Choo, “Scaling up personalized image aesthetic assessment via task vector customization,” in European Conference on Computer Vision. Springer, 2024, pp. 323–339

    2024

  17. [25]

    Uniaa: A unified multi-modal image aesthetic assessment baseline and benchmark,

    Z. Zhou, Q. Wang, B. Lin, Y . Su, R. Chen, X. Tao, A. Zheng, L. Yuan, P. Wan, and D. Zhang, “Uniaa: A unified multi-modal image aesthetic assessment baseline and benchmark,”arXiv preprint arXiv:2404.09619, 2024

    2024 arXiv

  18. [26]

    Personalized image aesthetics,

    J. Ren, X. Shen, Z. Lin, R. Mech, and D. J. Foran, “Personalized image aesthetics,” inThe IEEE International Conference on Computer Vision (ICCV), Oct 2017

    2017

  19. [27]

    Perceptual quality assessment of smartphone photography,

    Y . Fang, H. Zhu, Y . Zeng, K. Ma, and Z. Wang, “Perceptual quality assessment of smartphone photography,” inIEEE Conference on Computer Vision and Pattern Recognition, 2020, pp. 3677–3686

    2020

  20. [28]

    Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment,

    V . Hosu, H. Lin, T. Sziranyi, and D. Saupe, “Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment,”IEEE Transactions on Image Processing, vol. 29, p. 4041–4056, 2020. [Online]. Available: http://dx.doi.org/10.1109/TIP.2020.2967829

    2020

  21. [29]

    Grids: Grouped multiple-degradation restoration with image degradation similarity,

    S. Cao, Y . Liu, W. Zhang, Y . Qiao, and C. Dong, “Grids: Grouped multiple-degradation restoration with image degradation similarity,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 70–87

    2024

  22. [30]

    Diffvsr: Enhancing real- world video super-resolution with diffusion models for advanced visual quality and temporal consistency,

    X. Li, Y . Liu, S. Cao, Z. Chen, S. Zhuang, X. Chen, Y . He, Y . Wang, and Y . Qiao, “Diffvsr: Enhancing real- world video super-resolution with diffusion models for advanced visual quality and temporal consistency,” arXiv e-prints, pp. arXiv–2501, 2025

    2025

  23. [31]

    Dualx-vsr: Dual axial spatial ×temporal transformer for real-world video super-resolution without motion compensation,

    S. Cao, Y . Liu, X. Li, Y . Gao, Y . Zhou, and C. Dong, “Dualx-vsr: Dual axial spatial ×temporal transformer for real-world video super-resolution without motion compensation,” 2025. [Online]. Available: https://arxiv.org/abs/2506.04830

    2025 arXiv

  24. [32]

    Language models are unsupervised multitask learners,

    A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, I. Sutskeveret al., “Language models are unsupervised multitask learners,”OpenAI blog, vol. 1, no. 8, p. 9, 2019

    2019

  25. [33]

    Next token is enough: Realistic image quality and aesthetic scoring with multimodal large language model,

    M. Li, R. Wang, L. Sun, Y . Bai, and X. Chu, “Next token is enough: Realistic image quality and aesthetic scoring with multimodal large language model,” 2025. [Online]. Available: https://arxiv.org/abs/2503.06141

    2025 arXiv

  26. [34]

    Sgdr: Stochastic gradient descent with warm restarts,

    I. Loshchilov and F. Hutter, “Sgdr: Stochastic gradient descent with warm restarts,” inICLR, 2017

    2017

  27. [35]

    Musiq: Multi-scale image quality transformer,

    J. Ke, Q. Wang, Y . Wang, P. Milanfar, and F. Yang, “Musiq: Multi-scale image quality transformer,” 2021. [Online]. Available: https://arxiv.org/abs/2108.05997

    2021 arXiv

  28. [36]

    Vila: On pre-training for visual language models,

    J. Lin, H. Yin, W. Ping, Y . Lu, P. Molchanov, A. Tao, H. Mao, J. Kautz, M. Shoeybi, and S. Han, “Vila: On pre-training for visual language models,” 2023

    2023

  29. [37]

    mplug-owl2: Revolutionizing multi-modal large language model with modality collaboration,

    Q. Ye, H. Xu, J. Ye, M. Yan, A. Hu, H. Liu, Q. Qian, J. Zhang, F. Huang, and J. Zhou, “mplug-owl2: Revolutionizing multi-modal large language model with modality collaboration,” 2023. [Online]. Available: https://arxiv.org/abs/2311.04257 11

    2023 arXiv

  30. [38]

    Sharegpt4v: Improving large multi-modal models with better captions,

    L. Chen, J. Li, X. Dong, P. Zhang, C. He, J. Wang, F. Zhao, and D. Lin, “Sharegpt4v: Improving large multi-modal models with better captions,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 370–387

    2024

  31. [39]

    Q-instruct: Improving low-level visual abilities for multi-modality foundation models,

    H. Wu, Z. Zhang, E. Zhang, C. Chen, L. Liao, A. Wang, K. Xu, C. Li, J. Hou, G. Zhai, G. Xue, W. Sun, Q. Yan, and W. Lin, “Q-instruct: Improving low-level visual abilities for multi-modality foundation models,” 2023

    2023

  32. [40]

    Scaling up personalized image aesthetic assessment via task vector customization,

    J. Yun and J. Choo, “Scaling up personalized image aesthetic assessment via task vector customization,”

  33. [41]

    Methodology for the subjective assessment of the quality of television pictures,

    R. BT, “Methodology for the subjective assessment of the quality of television pictures,”International Telecommunication Union, vol. 4, p. 19, 2002. 12 Appendix Contents 1 Introduction 2 2 Related Work 3 2.1 Multi-modality Large Language Models . . . . . . . . . . . . . . . . ...

    2002

  34. [2023]

    Available: https://arxiv.org/abs/2310.17887

    [Online]. Available: https://arxiv.org/abs/2310.17887

    arXiv

  35. [2024]

    Available: https://arxiv.org/abs/2407.07176

    [Online]. Available: https://arxiv.org/abs/2407.07176

    arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.