ELIQ: Label-Free Quality Assessment for Evolving AI Images
ELIQ has been developed by researchers as a label-free system designed to evaluate the quality of images generated by AI, particularly as generative models advance. This framework emphasizes visual quality and the alignment between prompts and images, autonomously creating both positive and negative pairs to address traditional and AIGC-specific distortions without the need for human input. It modifies a pre-trained multimodal model into a quality-aware evaluator through instruction tuning, utilizing a lightweight gated fusion and a Quality Query Transformer for predicting two-dimensional quality. Experimental results indicate that ELIQ surpasses current label-free techniques across various benchmarks and effectively adapts to new distortions, tackling the issue of evolving quality standards in text-to-image models.
Key facts
- ELIQ is a label-free framework for quality assessment of evolving AI-generated images.
- It focuses on visual quality and prompt-image alignment.
- It automatically constructs positive and aspect-specific negative pairs.
- Pairs cover conventional distortions and AIGC-specific distortion modes.
- ELIQ adapts a pre-trained multimodal model via instruction tuning.
- It uses lightweight gated fusion and a Quality Query Transformer.
- Experiments show it outperforms existing label-free methods.
- The framework generalizes to new distortions without human annotations.
Entities
Institutions
- arXiv