Research map / Vision-language and multimodal / Human and embodied modeling

Rethinking Personalized Aesthetics Assessment: Employing Physique Aesthetics Assessment as An Exemplification

CVPR 2025 · Vision-language and multimodal › Human and embodied modeling · approach: Multimodal and affective interaction modeling · 2 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.

Proposes PAA+, a three-stage personalized aesthetics paradigm (pre-training, fine-tuning, continual learning), validated on physique aesthetics via a 50K-image PhysiqueAA50K dataset and the PhysiqueFrame model.

Problem

Personalized aesthetics assessment (PAA) relies on pre-training with generic aesthetics data whose collective preferences conflict with individual tastes, uses surveys that ignore object characteristics, and underuses multimodal user feedback during usage.

Approach

Replace the GAA pre-training with personality-specific pre-training (16 MBTI types), add a continual-learning stage that updates the user model from multimodal feedback, and use a 3D mesh/posture-based physique network with an LLM-based preference extraction network.

Key results

• PhysiqueFrame achieves appearance SRCC 0.699 and LCC 0.724 on User-ISFJ, outperforming the best baseline EAT (0.628 / 0.650).
• Personality-matched pre-training beats GAA pre-training across three users (e.g., User-ISFJ posture SRCC 0.664 vs 0.652 GAA-pretrained baseline comparisons in Table 1).
• User satisfaction rises from 75.3% to 84.5% for User-ISFJ with personality-matched pre-training and PhysiqueFrame, and continual feedback increases satisfaction over 4 update epochs (Fig. 7).

Measured effects reported in the paper

Surprising findings

• GAA collective preferences can violate transitivity (voting paradox), so GAA pre-training may introduce conflicts for personalized prediction.

Limitations

• Scoring alone is an insufficient means of assessment, as the authors acknowledge.
• Continual learning validated only on 3 users and 400 images over 4 epochs.
• Preference extraction relies on ChatGPT-generated sentiment labels.

Experimental setup

Datasets: PAA-16-personality PAA-3-User

Benchmarks: PhysiqueAA50K

Metrics: SRCC LCC binary classification accuracy user satisfaction rate

Baselines: NIMA BIAA HGCN TANet MaxViT TCFormer EAT Q-Align

Builds on: LLaVA LoRA Swin Transformer V2 GradCAM Chain-of-thought prompting

Closest papers by meaning

• Advancing Comprehensive Aesthetic Insight with Multi-Scale Text-Guided Self-Supervised Learning (AAAI 2025 · similarity 0.85)
• ArtiMuse: Fine-Grained Image Aesthetics Assessment with Joint Scoring and Expert-Level Understanding (CVPR 2026 · similarity 0.85)
• ELTA: An Enhancer against Long-Tail for Aesthetics-oriented Models (ICML 2024 · similarity 0.83)
• Bridging Cognitive Gap: Hierarchical Description Learning for Artistic Image Aesthetics Assessment (AAAI 2026 · similarity 0.83)
• Unlocking the Essence of Beauty: Advanced Aesthetic Reasoning with Relative-Absolute Policy Optimization (ICLR 2026 · similarity 0.82)
• Can Vision–Language Models Assess Graphic Design Aesthetics? A Benchmark, Evaluation, and Dataset Perspective. (ICLR 2026 · similarity 0.82)