ArtiMuse: Multimodal IAA Framework

• ArtiMuse is a multimodal LLM-based framework for image aesthetics assessment that delivers both continuous scores and detailed eight-dimensional critiques.
• It integrates visual and language models using a Token-As-Score strategy and gated cross-attention, ensuring balanced and interpretable multimodal analysis.
• ArtiMuse-10K, an expert-curated dataset of 10K images across diverse domains, supports robust evaluation of both quantitative scores and nuanced critic commentary.

Searching arXiv for the specified paper and closely related names to ground the article. ArtiMuse is a multimodal LLM–based framework for Image Aesthetics Assessment (IAA) that jointly predicts a continuous “expert” aesthetics score and generates a structured, eight-dimensional, fine-grained textual critique for arbitrary images, including photographs, paintings, AIGC outputs, 3D renderings, and graphic designs. It is introduced together with ArtiMuse-10K, an expert-curated dataset of 10,000 images spanning 5 main categories and 15 subcategories, each annotated with eight-dimensional attribute analysis and a holistic score. The system is positioned against prior MLLM-based IAA methods that are described as suffering from modality bias, in the sense of being score-only or text-only, and as lacking fine-grained attribute decomposition (Cao et al., 19 Jul 2025).

1. Scope, task definition, and naming

ArtiMuse addresses a specific formulation of IAA in which quantitative assessment and professional interpretation are treated as a single multimodal generation problem. The target output is not merely a scalar prediction, nor merely free-form commentary, but a paired assessment consisting of a continuous score in the range $[0,100]$ and a structured critique decomposed into eight aesthetic dimensions. This design reflects practical requirements arising from educational applications, artistic creation, and AIGC technologies, where downstream use often depends on both calibrated scoring and interpretable reasoning (Cao et al., 19 Jul 2025).

Within the paper’s framing, the central limitation of earlier MLLM-based IAA is modality asymmetry: some methods provide scores without expert-readable justification, while others produce text without robust fine-grained scoring. ArtiMuse is proposed as a joint alternative, with “Joint Scoring and Expert-Level Understanding” as its defining capability. A plausible implication is that the framework is intended not only for benchmark optimization, but also for analytic workflows in which textual explanation is part of the assessment object rather than an auxiliary artifact.

The name should be distinguished from similarly titled but substantively different resources. “ArtMUSE” is a Chinese-language multimodal creative-writing corpus associated with FlexMUSE and the task of multimodal creative writing (Chen et al., 22 Aug 2025), while “ArtistMus” is an artist-centric benchmark for retrieval-augmented music question answering paired with MusWikiDB (Kwon et al., 5 Dec 2025). ArtiMuse, by contrast, is an IAA model and dataset centered on expert aesthetic judgment (Cao et al., 19 Jul 2025).

2. Core architecture and joint prediction mechanism

The visual backbone is InternViT-300M-448px-V2.5, which tokenizes a $448\times448$ image $I$ into a sequence of vision tokens $V = E_v(I)\in\mathbb{R}^{N\times d}$ with $N\approx49$. The LLM is Qwen2.5-7B. Fusion is implemented by inserting a gated cross-attention module into each transformer block, attending from language hidden states to vision tokens. At layer $\ell$, the paper specifies

$$Q^\ell = H_t^\ell W_Q,\qquad K^\ell = V W_k,\qquad V^\ell = V W_v,$$

$$A^\ell = \mathrm{softmax}\!\left(\frac{Q^\ell K^{\ell T}}{\sqrt{d}}\right),\qquad O^\ell = A^\ell V^\ell,$$

with $O^\ell$ added back into the language stream to obtain the fused representation $H_t^{\prime \ell}$ (Cao et al., 19 Jul 2025).

A distinctive component is the “Token-As-Score” strategy. Instead of asking the LLM to generate a numeral directly or discretizing scores into coarse bins, the model reserves 101 special existing tokens $448\times448$0, each mapped to an integer score from 0 to 100. During training, the overall score $448\times448$1 is replaced by the single token $448\times448$2. At inference, the model outputs a distribution $448\times448$3, and the final continuous prediction is decoded by expectation:

$448\times448$4

This mechanism allows continuous score estimation to remain inside the autoregressive token-generation interface of the LLM rather than requiring a separate regression head (Cao et al., 19 Jul 2025).

The learning objective is a unified GPT-style cross-entropy over the full output sequence. If $448\times448$5 are target tokens containing both the eight attribute-level sentences and, in fine-tuning, exactly one score token, then

$448\times448$6

The paper describes this as jointly realizing text-generation loss and score-prediction loss within a single objective. In implementation, only the LLM is LoRA-adapted during both text pretraining and score fine-tuning, preserving most of the base model’s parameters (Cao et al., 19 Jul 2025).

3. Eight-dimensional expert-level understanding

ArtiMuse structures aesthetic analysis into eight independent sub-sentences, one for each attribute: Composition, Visual Elements, Technical Execution, Originality, Theme, Emotion, Overall Gestalt, and Comprehensive Evaluation (Cao et al., 19 Jul 2025). In the dataset description, the corresponding axes are phrased as Composition & Design; Visual Elements & Structure; Technical Execution; Originality & Creativity; Theme & Communication; Emotion & Viewer Response; Overall Gestalt; and Comprehensive Evaluation. The model is trained via prompt templates to produce exactly one paragraph for each attribute.

This decomposition is central to the system’s claim of “expert-level understanding.” Rather than treating aesthetic judgment as an opaque scalar, it externalizes distinct dimensions of evaluation that are conventionally entangled in end-to-end score prediction. The paper further argues that the cross-attention layers can thereby learn region-to-attribute correspondences, such as strong diagonals for “Composition” or color patches for “Visual Elements” (Cao et al., 19 Jul 2025). This suggests an implicit alignment between visual substructure and discourse structure, although the paper does not formalize that alignment beyond the architectural mechanism and prompt design.

The generated text is intended to be professionally toned rather than conversational. An example given for “Composition & Design” discusses rule-of-thirds placement, diagonal leading lines, negative space imbalance, and possible compositional counterpoint. The model can also embed simple LaTeX when an explanation invokes formulaic terms; the paper gives the Michelson contrast expression $448\times448$7 as an example of such output behavior (Cao et al., 19 Jul 2025).

A common misconception in aesthetic modeling is that explainability can be recovered after the fact from a score-only predictor. ArtiMuse rejects that separation by making textual critique a first-class supervised output. Another misconception is that textual critique alone suffices for IAA; the model’s paired design instead treats calibrated scoring and structured interpretation as mutually constraining outputs.

4. ArtiMuse-10K dataset

ArtiMuse-10K is described as the first expert-curated image aesthetic dataset comprising 10,000 images with both eight-dimensional attribute analysis and a holistic score. Images were chosen by a panel of art-and-photography experts with 3–30+ years’ experience. The collection spans five broad domains and 15 fine-grained subcategories (Cao et al., 19 Jul 2025).

Broad domain	Subcategories or examples
Photography	daily scenes, portrait, architecture, movie stills
Painting & Calligraphy	Chinese ink, sketch, children’s art, digital art, calligraphy
AIGC outputs	Stable Diffusion, Dreamlike, FLUX, etc.
3D Design	product, sculpture
Graphic Design	posters, logos

Each image carries prose annotations along the eight aesthetic dimensions and a global score in $448\times448$8. The paper states that source-dataset scores from heterogeneous ranges, including AVA’s 1–10 distributions and PARA’s 1–100, are normalized into the uniform 0–100 range before tokenization (Cao et al., 19 Jul 2025). This normalization is important because it permits the Token-As-Score mechanism to operate over a shared target scale across datasets.

The dataset’s composition signals a deliberate broadening of IAA beyond conventional photography benchmarks. Photography remains one domain, but the inclusion of painting, calligraphy, AIGC, 3D design, and graphic design makes the benchmark explicitly cross-medium. A plausible implication is that ArtiMuse-10K is designed to test whether aesthetic judgment models can generalize across media where evaluative criteria overlap only partially, such as composition and color balance on one hand, and medium-specific craft criteria on the other.

5. Training regimen and implementation choices

Training follows a two-stage regimen. The first stage is text pretraining on approximately 360,000 image-caption pairs, combining ArtiMuse-10K’s 10K examples with high-quality subsets of APDDv2, SPAQ, KonIQ, Impressions, AVA, TAD66K, PARA, and FLICKR-AES. This stage uses LoRA rank $448\times448$9, learning rate $I$0, batch size $I$1, one epoch, and cosine decay. The second stage is score fine-tuning on each target IAA dataset individually, with LoRA rank $I$2, $I$3, batch size $I$4, and 2 epochs (Cao et al., 19 Jul 2025).

All experiments are conducted on $I$5 NVIDIA A100 80 GB. The reported runtime is approximately 5 hours for pretraining and 10 minutes to 4 hours for each fine-tuning run, depending on dataset size. These details indicate that the proposed approach is intended to be practically trainable on contemporary multi-GPU infrastructure rather than requiring full-parameter adaptation of the base MLLM.

A specific implementation choice is the fixed-resolution strategy. Unlike InternVL’s tiling approach, all inputs are resized to $I$6 pixels in order to preserve global compositional cues. The paper reports that this yields approximately 0.3 points SRCC/PLCC gain on scored benchmarks and more than $I$7 throughput (Cao et al., 19 Jul 2025). The rationale is consistent with the aesthetic target: compositional assessment often depends on spatial globality, and aggressive tiling can perturb exactly those scene-level relationships that drive composition and gestalt judgments.

6. Evaluation, benchmark results, and ablations

Evaluation is divided into structural aesthetic analysis and aesthetic scoring. For structural analysis, Gemini-2.0-flash is asked to choose the best of four models—AesExpert, Qwen-2.5-VL, InternVL3-8B, and ArtiMuse—for each of the eight attributes, using human critic text as reference. ArtiMuse is selected 71.1% of the time, versus 14.5% for the next best model. In a blind human study over 20 images and 8 attributes, volunteers prefer ArtiMuse’s critique 67.8% of the time, compared with 19.2% for InternVL3 and 11.5% for Qwen-2.5 (Cao et al., 19 Jul 2025).

For scoring, the paper reports SRCC and PLCC on multiple datasets:

Dataset	SRCC	PLCC
AVA	0.827	0.826
PARA	0.936	0.958
TAD66K	0.510	0.543
FLICKR-AES	0.814	0.837
ArtiMuse-10K (zero-shot)	0.614	0.627

The reported interpretations are dataset-specific: on AVA, ArtiMuse is second-best in SRCC and best in PLCC; on PARA and FLICKR-AES it is best on both metrics; on TAD66K it is second in SRCC and best in PLCC; and on ArtiMuse-10K zero-shot it is best on both metrics (Cao et al., 19 Jul 2025).

The paper also reports zero-shot generalization: when trained only on AVA, ArtiMuse still outperforms Q-Align on four unseen datasets by +0.03 to +0.06 PLCC. This is presented as evidence that the joint text-plus-score formulation does not merely memorize dataset-specific score distributions, but transfers across evaluation settings (Cao et al., 19 Jul 2025).

The ablation studies clarify which components drive performance. Removing expert images with score captions reduces AVA performance to 0.621/0.627. Replacing LoRA plus two-stage training with full LLM fine-tuning reduces AVA from 0.827/0.826 to 0.816/0.814. Using joint text+score training rather than the two-stage strategy yields 0.821/0.820. For score tokenization granularity, 100 tokens gives 0.827/0.826, outperforming 5 levels at 0.820/0.818; 25 and 50 tokens lie between them, while 200 tokens is worse than 100 tokens (Cao et al., 19 Jul 2025). These results indicate that both expert-curated supervisory text and the specific tokenized scoring design materially affect performance.

7. Limitations, interpretation, and future directions

The stated limitation is that ArtiMuse currently analyzes and evaluates but does not yet suggest concrete “how-to” enhancements, such as “increase exposure by +0.5 EV” or “introduce a warm color accent” (Cao et al., 19 Jul 2025). In other words, its output is diagnostic rather than prescriptive. This matters because many practical aesthetic workflows, especially in design iteration and computational photography, require actionable modification proposals in addition to critique.

Future work is described as a third training stage in which the model, conditioned on the eight-dimensional critique, generates actionable editing instructions or aesthetic-enhancement prompts, potentially via a small toolkit of differentiable augmentation operators (Cao et al., 19 Jul 2025). This suggests a transition from assessment to intervention: the same structured representation that supports expert-style critique could serve as an intermediate control space for image refinement.

More broadly, ArtiMuse represents a move in IAA from scalar prediction toward joint semantic assessment. Its dataset design, multimodal architecture, Token-As-Score formulation, and two-stage supervision strategy all reflect the premise that aesthetic judgment is neither reducible to a single number nor adequately captured by unconstrained descriptive text alone. Within that framing, the system’s contribution is not simply a benchmark improvement, but a redefinition of IAA as a multimodal expert-critique problem with quantitative and discursive outputs integrated in one model (Cao et al., 19 Jul 2025).