EvoQuality - ByteDance's Open-Source Image Quality Assessment Model

What is EvoQuality

EvoQuality is a self-evolving vision-language model framework jointly launched by ByteDance and City University of Hong Kong, focusing on no-reference image quality assessment. The model is built on Qwen2.5-VL-7B and requires no human-annotated quality scores or distortion labels at all. It generates pseudo-ranking labels through the model's own pairwise comparisons and majority voting, and then uses the GRPO reinforcement learning algorithm to achieve multi-round iterative self-evolution.

Main Features of EvoQuality

Single-Image Quality Scoring: Outputs a continuous quality score for a single image, supports multiple distortion types, including real distortions, synthetic distortions, and AI-generated distortions.
Image Pair Quality Comparison: Determines the relative quality superiority or inferiority of two images through pairwise comparison, and generates interpretable quality description text.
Self-Evolving Iterative Training: In the offline stage, high-confidence pseudo-labels are generated through majority voting; in the online stage, the policy is optimized through GRPO, forming a closed loop of self-improvement.

Technical Principles of EvoQuality

Offline Pseudo-Label Generation: Multiple queries are performed on unlabeled image pairs, letting the current VLM judge "which image has better quality." Through Pairwise Majority Voting, a relative quality consensus is established, pseudo-ranking labels are generated, and human MOS annotations are completely replaced. The Thurstone Case V psychometric model is used to transform discrete comparison results into continuous quality score distributions, generating an optimizable fidelity reward signal.
Online Policy Evolution: The GRPO algorithm is used to transform pseudo-labels into reward signals and update the VLM policy. The advantage function is estimated through the relative rewards of samples within a group, greatly reducing training memory and computational overhead. The policy model generates multiple responses for the same batch of image pairs, and performs gradient updates based on the rewards calculated from the pseudo-labels.
Iterative Evolution Mechanism: Multiple rounds of iteration form positive feedback: improved model capability → generation of higher-quality pseudo-labels → further model evolution. Experiments show that zero-shot PLCC improves by an average of 31.8%.

How to Use EvoQuality

Environment Preparation: Install Python 3.8+ and dependency libraries such as PyTorch and Transformers, and ensure that a GPU environment is available.
Model Loading: Pull from Hugging Face ByteDance/EvoQuality the model weights and processor files to the local machine.
Single-Image Quality Scoring: Read the image to be evaluated, construct a prompt such as "Please rate the quality of this image from 0 to 100." and input it into the model. The model outputs a continuous quality score and structured text describing quality defects/strengths.
Image Pair Quality Comparison: Prepare two images to be compared, construct a prompt such as "Which image has better quality? Explain why." and input it into the model.
Batch Evaluation: Perform batch inference on an image dataset, and combine the majority voting mechanism from multiple queries to generate high-confidence pseudo-labels.
Self-Evolving Training (Advanced): Use the generated pseudo-labels to fine-tune the model through the GRPO algorithm, and start a new round of iterative evolution to continuously improve assessment accuracy.

Core Advantages of EvoQuality

Zero Annotation Cost: Completely requires no human subjective ratings or distortion labels; training signals can be generated solely through the model's own pairwise comparisons and majority voting.
Performance Surpasses Supervised Models: Surpasses current state-of-the-art supervised VLM-based IQA methods on 5 out of 7 IQA benchmarks, with an average zero-shot PLCC improvement of 31.8%.
Self-Evolving Closed-Loop Capability: Through multiple rounds of iteration, it forms a positive feedback loop of generating pseudo-labels → training the model → a stronger model → generating better labels, continuously breaking through the performance ceiling.
Strong Cross-Dataset Generalization: Naturally supports zero-shot cross-domain evaluation, without the need to realign the perceptual scale for a new dataset or retrain.

EvoQuality Project Address

HuggingFace Model Repository：https://huggingface.co/ByteDance/EvoQuality
arXiv Technical Paper：https://arxiv.org/pdf/2509.25787

Comparison of EvoQuality with Similar Competitors

Dimension	EvoQuality	VisualQuality-R1
Supervision Method	Fully self-supervised, zero manual annotation	Requires manual MOS annotations as ground truth
Core Algorithm	GRPO + pairwise majority voting pseudo-labels + multi-round self-evolution	GRPO + Thurstone model + continuous fidelity reward
Reward Source	Pseudo-ranking labels generated by the model itself (no external annotation required)	Continuous fidelity measure computed based on manual MOS
Model Foundation	Qwen2.5-VL-7B	Qwen2.5-VL-7B
Training Mechanism	Offline pseudo-label generation → online GRPO optimization → iterative closed-loop evolution	Single-round/limited-round RL training, dependent on fixed annotated datasets
Data Dependence	Requires only unannotated images, with zero data acquisition cost	Requires datasets with MOS such as KADID-10K, TID2013, and KonIQ-10k
Iteration Capability	Supports multiple rounds of bootstrapped iteration, with model and label quality mutually improving	No further evolution after training convergence, limited by the scale of annotated data
Interpretability	Outputs quality score + structured quality description text	Outputs quality score + reasoning process (thinking) + quality description
Cross-Dataset Training	Naturally supported, no need for perceptual scale realignment	Supports multi-dataset training, no need for scale realignment

Application Scenarios of EvoQuality

AIGC Content Quality Inspection: Automatically evaluates the perceptual quality of AI-generated content such as text-to-image, image super-resolution, and old photo restoration, replacing manual review and reducing the operating costs of content platforms.
Image Compression and Transmission Optimization: Evaluates image quality after compression in real time, finding the optimal balance between bandwidth and image quality, suitable for scenarios such as video streaming and cloud storage.
Mobile Photography Assistance: Analyzes the quality of the framing view in real time, guiding users to adjust focus, exposure, and composition parameters, or automatically selecting the best shot from a burst.
Video Quality Assessment: Extends single-frame evaluation to the temporal dimension, analyzing the impact of motion blur, inter-frame consistency, and bitrate fluctuations on perceived quality, for video encoding and transmission quality monitoring.