ByteDance/EvoQuality
EvoQuality is a 7 billion parameter self-evolving Vision-Language Model (VLM) developed by ByteDance for No-Reference Image Quality Assessment (NR-IQA), supporting both single-image quality scoring and pairwise comparison. Built on a Qwen2.5-VL-7B backbone, it iteratively refines its quality perception without human-annotated scores by generating pseudo-ranking labels via pairwise majority voting and optimizing through Group Relative Policy Optimization (GRPO). This model significantly improves PLCC and SRCC metrics over its backbone, demonstrating strong generalization across diverse distortion types and AI-generated content, making it suitable for research, evaluation, and pre-production quality assessment.
Loading preview...
EvoQuality: Self-Evolving VLM for Image Quality Assessment
EvoQuality is a 7 billion parameter Vision-Language Model (VLM) developed by ByteDance, specifically designed for No-Reference Image Quality Assessment (NR-IQA). Unlike traditional methods, EvoQuality operates without relying on human-annotated quality scores or distortion labels. It achieves its quality perception capabilities through a unique self-evolution process.
Key Capabilities
- Self-Evolving Quality Perception: Generates pseudo-ranking labels via pairwise majority voting and iteratively refines its understanding of image quality using Group Relative Policy Optimization (GRPO).
- Dual Assessment Modes: Supports both single-image quality scoring (rating from 1 to 5) and pairwise image quality comparison (ranking).
- Significant Performance Improvement: Achieves substantial gains over its Qwen2.5-VL-7B backbone, with PLCC improving from 0.615 to 0.770 (+31.8%) and SRCC from 0.570 to 0.726 (+33.7%).
- Strong Generalization: Demonstrates improved performance across diverse distortion types and AI-generated content, matching or surpassing several supervised VLM-IQA approaches.
- Zero-Shot Evaluation: Evaluated in a zero-shot setting, indicating robust performance without training on target test sets.
Good For
- Research and Evaluation: Ideal for NR-IQA studies, cross-dataset generalization comparisons, quality ranking/filtering, and generating auxiliary signals for data cleaning.
- Pre-production Assessment: Can serve as a perceptual quality proxy in early development stages, though it should be combined with business data and manual validation for critical decisions.
- Understanding Model Self-Improvement: Provides a practical example of a model that self-evolves its capabilities without direct human supervision on labels.
Paper: Self-Evolving Vision-Language Models for Image Quality Assessment via Voting and Ranking