Research map / Vision-language and multimodal / Multimodal LLMs

Bridging Cognitive Gap: Hierarchical Description Learning for Artistic Image Aesthetics Assessment

AAAI 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Interleaved image-text generation with MLLMs · 1 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.

Builds the RAD dataset of 70k hierarchical aesthetic descriptions generated by GPT-4o with DeepSeek-based consistency checking, and trains ArtQuant, an MLLM that jointly learns description generation and score prediction via a score-based distribution estimation loss.

Problem

Artistic image aesthetics assessment lacks multi-dimensional (perception, cognition, emotion) annotated descriptions, and existing visual or contrastive multimodal methods struggle with long-form text and continuous score prediction.

Approach

Generate a large hierarchical aesthetic description dataset with an LLM generator and LLM-based consistency filter, then train an MLLM with an auxiliary description-generation task (multi-task aesthetic training) followed by score-based distribution estimation that maps discrete token probabilities to continuous scores without requiring label variance.

Key results

• ArtQuant reaches SROCC 0.871 and PLCC 0.894 on APDD, versus 0.810/0.840 for ArtCLIP.
• On BAID, ArtQuant achieves SROCC 0.543 and PLCC 0.589, above PVAFE's 0.533/0.583.
• On VAPS, ArtQuant achieves SROCC 0.625 and PLCC 0.681, above AKA-Net's 0.579/0.638, with fewer training epochs.

Surprising findings

• Generalist MLLMs such as GPT-4o perform far below specialist models on BAID, which the authors attribute to its narrow score distribution.
• Chain-of-thought prompting did not help GPT-4o on BAID (SROCC 0.067 with CoT vs 0.122 without).

Limitations

• Descriptions are generated by GPT-4o and may contain hallucinations, which the authors acknowledge.
• Base MLLM architecture and size are not clearly specified in the provided text.
• Theoretical bounds rely on assumptions such as epsilon-approximate conditional independence that are hard to verify.

Experimental setup

Datasets: APDD BAID VAPS RAD

Benchmarks: APDD BAID VAPS

Metrics: SROCC PLCC

Baselines: GPT-4o Qwen-VL-Plus SAAN AANSPS ArtCLIP LITA EAMB-Net AKA-Net PVAFE

Builds on: DeQA ArtCLIP SAAN AKA-Net

Closest papers by meaning

• Unlocking the Essence of Beauty: Advanced Aesthetic Reasoning with Relative-Absolute Policy Optimization (ICLR 2026 · similarity 0.90)
• Advancing Comprehensive Aesthetic Insight with Multi-Scale Text-Guided Self-Supervised Learning (AAAI 2025 · similarity 0.90)
• ArtiMuse: Fine-Grained Image Aesthetics Assessment with Joint Scoring and Expert-Level Understanding (CVPR 2026 · similarity 0.90)
• Can Vision–Language Models Assess Graphic Design Aesthetics? A Benchmark, Evaluation, and Dataset Perspective. (ICLR 2026 · similarity 0.89)
• APDDv2: Aesthetics of Paintings and Drawings Dataset with Artist Labeled Scores and Comments (NeurIPS 2024 · similarity 0.88)
• Aligning Vision Models with Human Aesthetics in Retrieval: Benchmarks and Algorithms (NeurIPS 2024 · similarity 0.88)