Research map / Vision-language and multimodal / Multimodal LLMs
Regression over Classification: Assessing Image Aesthetics via Multimodal Large Language Models
Xingyuan Ma, Shuai He, Anlong Ming, Haobin Zhong, Huadong Ma
Read the paperCheck your idea against 137,217 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.
Adapts MLLMs for image aesthetics assessment by replacing classification-based score decoding with a regression-style approach, addressing score insensitivity from tokenization.
Problem
Applying multimodal large language models to image aesthetics assessment (IAA) suffers from score insensitivity due to tokenization and from a classification-based decoding mechanism that ignores the numeric ordering of scores.
Approach
Reframes MLLM-based aesthetic scoring as regression rather than classification over score tokens, so predicted scores respect numeric proximity; the abstract only partially states the mechanism.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add decoupled score tokens and regression (ablation) | mPLUG-Owl2 baseline · AVA | SRCC: 0.788 -> 0.825 | Improved |
| add decoupled score tokens and regression (ablation) | mPLUG-Owl2 baseline · AVA | PLCC: 0.786 -> 0.824 | Improved |
| add decoupled score tokens and regression (ablation) | mPLUG-Owl2 baseline · ICAA17K | SRCC: 0.688 -> 0.873 | Improved |
| add decoupled score tokens and regression (ablation) | mPLUG-Owl2 baseline · ICAA17K | PLCC: 0.707 -> 0.880 | Improved |
| add loss decoupling via position token (ablation) | ROC4MLLM regression · AVA | SRCC: 0.825 -> 0.833 | Improved |
| add loss decoupling via position token (ablation) | ROC4MLLM regression · ICAA17K | SRCC: 0.873 -> 0.894 | Improved |
| add loss decoupling via position token (ablation) | ROC4MLLM regression · ICAA17K | PLCC: 0.880 -> 0.903 | Improved |
| add ROC4MLLM to mPLUG-Owl2 | mPLUG-Owl2 · AVA (score only) | SRCC: 0.761 -> 0.831 | Improved |
| add ROC4MLLM to Qwen2-VL | Qwen2-VL · AVA (score only) | SRCC: 0.784 -> 0.831 | Improved |
| add ROC4MLLM to LLaVA 1.5 | LLaVA 1.5 · AVA (score only) | SRCC: 0.775 -> 0.793 | Improved |
| add comment generation to training | mPLUG-Owl2 · AVA | SRCC: 0.761 -> 0.758 (-0.3%) | Hurt |
| add comment generation to training | Q-Align · AVA | SRCC: 0.821 -> 0.819 (-0.2%) | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- Q-Align — Q-Align: Teaching LMMs for Visual Scoring via Discrete Text-Defined Levels (ICML 2024) · better on 9 of 9 shared results
- mPLUG-Owl2 — mPLUG-Owl2: Revolutionizing Multi-modal Large Language Model with Modality Collaboration (CVPR 2024) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- ArtiMuse: Fine-Grained Image Aesthetics Assessment with Joint Scoring and Expert-Level Understanding (CVPR 2026 · similarity 0.92)
- Advancing Comprehensive Aesthetic Insight with Multi-Scale Text-Guided Self-Supervised Learning (AAAI 2025 · similarity 0.92)
- Teaching Large Language Models to Regress Accurate Image Quality Scores Using Score Distribution (CVPR 2025 · similarity 0.89)
- Probabilistic Prompt Adaptation for Unified Image Aesthetics and Quality Assessment (CVPR 2026 · similarity 0.89)
- Bridging Cognitive Gap: Hierarchical Description Learning for Artistic Image Aesthetics Assessment (AAAI 2026 · similarity 0.89)
- Revisiting MLLM Based Image Quality Assessment: Errors and Remedy (AAAI 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
