ERNIE-Image-Aes: Robust Image Aesthetics Scoring with Balanced Category Generalization
[📄 Paper]
🌟 Highlights
ERNIE-Image-Aes is a 8B vision-language model for image aesthetic scoring, initialized from ArtiMuse and fine-tuned on a diverse, professionally annotated dataset. It substantially outperforms existing aesthetic predictors (LAION-AES, ArtiMuse, UniPercept) in generalization across diverse image categories.
Key advantages:
- Balanced predictions across photography, anime, design, everyday snapshots, and film photography
- No systematic bias toward specific image types (e.g., AI-generated content or black-and-white photos)
- Swiss-tournament based pairwise annotation for high-quality training labels
- Achieves 0.7445 SRCC and 0.7598 PLCC on ERIA-1K benchmark
🔍 Motivation
Off-the-shelf aesthetic predictors exhibit systematic biases:
| Model | Bias |
|---|---|
| LAION-Aesthetic | Disproportionately high scores for AI-generated/anime content |
| ArtiMuse | Overscores black-and-white photography and casual everyday snapshots |
| UniPercept | Strong preference for monochrome images; overscores casual snapshots |
ERNIE-Image-Aes addresses these failure modes through a purpose-built annotation pipeline with explicit category balance.
📊 Results on ERIA-1K Benchmark
| Model | SRCC | PLCC |
|---|---|---|
| LAION AES | 0.2944 | 0.3138 |
| ArtiMuse | 0.4277 | 0.4704 |
| UniPercept | 0.4533 | 0.4748 |
| ERNIE-Image-Aes | 0.7445 | 0.7598 |
Annotation Protocol:
- Pairwise Swiss-system tournament for stable and reproducible rankings
- Tier labels from 1 to 10
- Annotators recruited from professional backgrounds (Central Academy of Fine Arts, Sichuan Fine Arts Institute, Communication University of China, etc.)
- All annotators passed aesthetic calibration screening prior to participation
⚙️ Setup
Please follow the setup instructions in the ArtiMuse repository.
🙏 Acknowledgements
Our work builds upon ArtiMuse and InternVL-3. We sincerely thank the authors for their excellent contributions to the community.
✒️ Citation
If you find this work useful, please consider citing: