Research map / Vision-language and multimodal / Multimodal LLMs

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

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Interleaved image-text generation with MLLMs · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces ArtiMuse, an InternVL3-8B-based MLLM that jointly produces expert-style 8-attribute aesthetic analysis and continuous scores via a Token-As-Score decoding scheme, plus the expert-annotated ArtiMuse-10K dataset.

Problem

MLLM-based image aesthetics assessment lacks both accurate continuous scoring and fine-grained, expert-level attribute analysis; existing datasets are small, coarse, and mostly photographic.

Approach

Map 101 existing LLM vocabulary tokens to integer scores 0-100 and predict the final score as the probability-weighted expectation over those tokens, avoiding vocabulary expansion and discretization into coarse levels; train with two-stage LoRA (text pretraining then score fine-tuning) on a mix of expert-annotated and MLLM-annotated aesthetic captions.

Key results

• ArtiMuse reaches PLCC 0.958 on PARA and 0.627 on ArtiMuse-10K, over 0.05 PLCC above the next best model on those datasets.
• In LLM-judged attribute analysis, ArtiMuse wins 71.1% of attribute comparisons on average versus 14.3% for Qwen-2.5-VL-7B and 14.5% for InternVL3-8B.
• With LoRA and two-stage training, 100 score tokens achieve SRCC 0.827 / PLCC 0.826 on AVA, versus 0.816 / 0.814 with full fine-tuning.

Surprising findings

• Full fine-tuning of the LLM degrades scoring relative to LoRA with two-stage training, attributed to loss of pretrained aesthetic priors.
• Expanding the vocabulary with new score tokens performs poorly on ArtiMuse-10K, while reusing existing tokens works better.

Limitations

• Model only analyzes and does not provide professional aesthetic enhancement recommendations.
• Attribute analysis is judged mainly by an MLLM (Gemini-2.0-flash) rather than fully by humans.

Experimental setup

Datasets: ArtiMuse-10K AVA APDDv2 PARA Impressions TAD66K FLICKR-AES SPAQ KonIQ-10K

Benchmarks: AVA PARA TAD66K FLICKR-AES ArtiMuse-10K

Metrics: SRCC PLCC selection rate

Baselines: Q-Align AesExpert Qwen-2.5-VL-7B InternVL3-8B TANet AesMamba VILA UNIAA-LLaVA Next Token Is Enough PEAS

Builds on: Q-Align InternVL3

Closest papers by meaning

• Advancing Comprehensive Aesthetic Insight with Multi-Scale Text-Guided Self-Supervised Learning (AAAI 2025 · similarity 0.91)
• Bridging Cognitive Gap: Hierarchical Description Learning for Artistic Image Aesthetics Assessment (AAAI 2026 · similarity 0.90)
• Unlocking the Essence of Beauty: Advanced Aesthetic Reasoning with Relative-Absolute Policy Optimization (ICLR 2026 · similarity 0.88)
• ELTA: An Enhancer against Long-Tail for Aesthetics-oriented Models (ICML 2024 · similarity 0.87)
• AACP: Aesthetics Assessment of Children’s Paintings Based on Self-Supervised Learning (AAAI 2024 · similarity 0.87)
• Can Vision–Language Models Assess Graphic Design Aesthetics? A Benchmark, Evaluation, and Dataset Perspective. (ICLR 2026 · similarity 0.87)