EvoQuality - ByteDance's open-source image quality assessment model - AiBoss

What is EvoQuality?

EvoQuality is a self-evolving visual language model framework jointly developed 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 manually labeled quality scores or distortion labels. It generates pseudo-ranking labels through pairwise comparisons and majority voting, and then uses the GRPO reinforcement learning algorithm to achieve multi-round iterative self-evolution.

EvoQuality's main functions

Single image quality scoreIt outputs continuous quality scores for a single image and supports multiple distortion types, including real distortion, synthetic distortion, and AI-generated distortion.
Image quality comparisonIt determines the relative quality of two images by comparing them in pairs and generates interpretable quality description text.
Self-evolutionary iterative trainingIn the offline phase, high-confidence pseudo-labels are generated through majority voting, and in the online phase, a GRPO optimization strategy is used to form a closed-loop self-improvement mechanism.

EvoQuality's Technical Principles

Offline pseudo-tag generationThe algorithm performs multiple queries on unlabeled image pairs, allowing the current Virtual Model (VLM) to determine "which image is of better quality." A relative quality consensus is established through pairwise majority voting, generating pseudo-ranking labels that completely replace manual MOS (Method of Motion) labeling. The Thurstone Case V psychometric model transforms discrete comparison results into a continuous quality score distribution, generating an optimizable fidelity reward signal.
Online Strategy EvolutionThe GRPO algorithm is used to convert pseudo-labels into reward signals to update the VLM policy. The advantage function is estimated by the relative reward of samples within the group, significantly reducing training memory and computational overhead. The policy model generates multiple responses for the same batch of image pairs, and updates the policy based on the gradient of the reward calculated from the pseudo-labels.
Iterative evolution mechanismMultiple iterations create positive feedback, improving model capabilities → generating higher-quality pseudo-labels → further model evolution. Experiments show that zero-sample PLCC achieves an average improvement of 31.8%.

How to use EvoQuality

Environmental preparationInstall Python 3.8+ and related libraries such as PyTorch and Transformers, and ensure that the GPU environment is available.
Model loadingPull from Hugging Face ByteDance/EvoQuality The model weights and processor files are saved locally.
Single image quality scoreThe system reads the image to be evaluated, constructs a prompt such as "Please rate the quality of this image from 0 to 100," and inputs it into the model. The model outputs a continuous quality score and structured text describing quality defects/advantages.
Image quality comparisonPrepare two images to be compared, construct prompts such as "Which image has better quality? Explain why." and input them into the model.
Batch evaluationBatch inference is performed on the image dataset, and a majority voting mechanism based on multiple queries is used to generate high-confidence pseudo-labels.
Self-evolutionary training (advanced)The generated pseudo-labels are used to fine-tune the model using the GRPO algorithm, initiating a new round of iterative evolution to continuously improve the evaluation accuracy.

EvoQuality's core advantages

Zero labeling costIt requires no subjective human scoring or distorted labels; training signals can be generated solely through pairwise comparisons and majority voting within the model itself.
Performance surpasses supervised modelsIt outperforms state-of-the-art supervised VLM-based IQA methods on 5 out of 7 IQA benchmarks, with an average improvement of 31.8% in zero-sample PLCC.
Self-evolutionary closed-loop capabilityThrough multiple iterations, a positive feedback loop is formed: generating pseudo-labels → training the model → making the model stronger → generating better labels, continuously breaking through the performance ceiling.
Strong generalization across datasetsIt naturally supports zero-shot cross-domain evaluation without requiring realignment of the perception scale or retraining for new datasets.

EvoQuality's project address

HuggingFace model libraryhttps://huggingface.co/ByteDance/EvoQuality
arXiv technical paper: https://arxiv.org/pdf/2509.25787

Comparison of EvoQuality's similar products

Dimension	EvoQuality	VisualQuality-R1
Supervision methods	Fully self-supervised, zero manual annotation	Manual MOS annotation is required as ground truth.
Core Algorithm	GRPO + Paired majority voting pseudo-labels + Multi-round self-evolution	GRPO + Thurstone model + continuous fidelity bonus
Reward Source	The model generates its own pseudo-ranking labels (no external annotation required).	Continuous fidelity measure based on artificial MOS calculations
Model Basics	Qwen2.5-VL-7B	Qwen2.5-VL-7B
Training mechanism	Offline pseudo-tag generation → Online GRPO optimization → Iterative closed-loop evolution	Single-round/limited-round RL training, relying on a fixed labeled dataset
Data dependency	Data acquisition is cost-free, requiring only unlabeled images.	Requires datasets with MOS such as KADID-10K, TID2013, and KonIQ-10k.
Iteration capability	Supports multiple rounds of bootstrapping iterations, allowing for mutual improvement in model and label quality.	Once training converges, it stops evolving, limited by the size of the labeled data.
Explainability	Output quality score + structured quality description text	Output quality score + reasoning process + quality description
Training across datasets	Natural support, no need for perceived scale realignment	Supports training on multiple datasets without the need for scale realignment.

Application Scenarios of EvoQuality

AIGC Content Quality InspectionAutomatically assess the perceived quality of AI-generated content such as text-to-image editing, image super-resolution, and old photo restoration, replacing manual review and reducing the operating costs of content platforms.
Image compression and transmission optimizationIt can evaluate the quality of compressed images in real time and find the optimal balance between bandwidth and image quality, making it suitable for scenarios such as video streaming and cloud storage.
Mobile phone photography aidsIt analyzes the quality of the viewfinder image in real time, guiding users to adjust focus, exposure, and composition parameters, or automatically selects the best image from continuous shooting.
Video quality assessmentExtending single-frame evaluation to the temporal dimension, this study analyzes the impact of dynamic blur, inter-frame consistency, and bitrate fluctuations on viewing experience, and is used for video encoding and transmission quality monitoring.