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

Haobin Zhong, Shuai He, Anlong Ming, Huadong MaCVPR 2025 · DOI

Introduction

Beauty is notoriously subjective, yet most AI models for aesthetic assessment are trained on "collective" data that often contradicts individual taste. This paper introduces PAA+, a new three-stage paradigm—pre-training, fine-tuning, and continual learning—that uses personality-matched models and multimodal feedback to capture unique human preferences. The authors apply this framework to the complex domain of human physique, demonstrating how specialized architectures and personalized priors can overcome the limitations of generic models.

The Voting Paradox: Why Collective Data Fails

A critical flaw in current Personalized Aesthetics Assessment (PAA) is the reliance on Generic Aesthetics Assessment (GAA) datasets for pre-training. GAA reflects collective preferences, which often suffer from non-transitivity—a phenomenon known as the "Voting Paradox."

The Voting Paradox: Individual preferences are transitive, but collective preferences can become cyclic (e.g., A > B > C > A), making them a poor foundation for learning individual taste.

As illustrated in the diagram above, while individual annotators have consistent, transitive rankings, the aggregated "crowd" preference can result in a cycle (A > B, B > C, but C > A). This mathematical inconsistency makes GAA a noisy and often irrational foundation for models intended to predict specific individual preferences.

The PAA+ Paradigm

The researchers propose the PAA+ paradigm, which expands the traditional "pre-train and fine-tune" workflow into a more robust three-stage process.

Comparison between the standard PAA approach (left) and the proposed PAA+ paradigm (right), highlighting the shift to personality-matched bases and continual learning.
  • Stage 1: Personality-Matched Pre-training: Instead of one generic model, the system develops 16 distinct models corresponding to MBTI personality types. A Human-AI collaborative loop (RLHF with LLaVA++) is used to generate and refine "pseudo-labels" for these models.
  • Stage 2: Fine-tuning: A user selects the personality-matched model that best fits their profile and fine-tunes it on their specific data.
  • Stage 3: Continual Learning: The model continuously evolves by processing multimodal user feedback—such as text comments and revised scores—during active use.

Implementation: PhysiqueFrame and PhysiqueAA50K

To implement this paradigm, the authors developed PhysiqueFrame, which utilizes two specialized networks:

  • PANet (Physique Analysis Network): Extracts 3D-aware features using a Mesh Perceiving Module to analyze body shape and a Posture Analyzing Module to evaluate skeletal keypoints.
  • PENet (Preference Extraction Network): Uses a Large Language Model (LLaVA++) and Chain-of-Thought (CoT) reasoning to translate diverse user feedback into quantitative preference scores.

The study is supported by the PhysiqueAA50K dataset, which includes over 50,000 images. This benchmark features PAA-16-Personality (40,000 images annotated across 16 MBTI types) and PAA-3-User (a validation set for testing fine-tuning and continual learning).

Key Results

  • Superior Initialization: Models pre-trained on personality-specific data consistently outperformed those pre-trained on generic GAA data across all metrics (SRCC, LCC, and Accuracy).
  • Continual Improvement: In the "Continual Learning" stage, user satisfaction rates climbed steadily as the model processed more feedback, reaching over 84% satisfaction for test users.
  • State-of-the-Art Performance: PhysiqueFrame outperformed eight representative baseline models, proving that its specialized 3D-aware architecture captures physique-related features more effectively than general aesthetic models.

Conclusion

The PAA+ paradigm represents a significant shift in aesthetic assessment, moving from noisy collective averages to personalized, evolving models. By leveraging personality-matched priors and multimodal feedback loops, AI can better align with the subjective nature of human beauty. Future work will explore more granular ways to initialize personalized models beyond the 16 MBTI types.

Summary updated

BibTeX
@inproceedings{zhong2025rethinking,
  author = {Haobin Zhong and Shuai He and Anlong Ming and Huadong Ma},
  title = {{Rethinking Personalized Aesthetics Assessment: Employing Physique Aesthetics Assessment as An Exemplification}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2025},
  doi = {10.1109/cvpr52734.2025.00279},
}
Figure