Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Q-Refine: A Perceptual Quality Refiner for AI-Generated Image

Chunyi Li, Haoning Wu, Zicheng Zhang, Hongkun Hao, Kaiwei Zhang, Lei Bai, Xiaohong Liu, Xiongkuo Min, Weisi Lin, Guangtao Zhai

TL;DR

The paper tackles inconsistent perceptual quality in AI-generated images by introducing Q-Refine, a quality-aware refiner guided by an Image Quality Assessment map. It deploys three adaptive pipelines—Gaussian noise, patch-based inpainting, and global enhancement—to selectively refine low-, medium-, and high-quality regions, respectively. Across multiple AIGI quality datasets and T2I models, Q-Refine demonstrates broad improvements in both fidelity and aesthetic quality, without degrading already high-quality regions. This perceptual, region-specific refinement approach broadens the practical utility of T2I systems by delivering consistent, quality-aware improvements.

Abstract

With the rapid evolution of the Text-to-Image (T2I) model in recent years, their unsatisfactory generation result has become a challenge. However, uniformly refining AI-Generated Images (AIGIs) of different qualities not only limited optimization capabilities for low-quality AIGIs but also brought negative optimization to high-quality AIGIs. To address this issue, a quality-award refiner named Q-Refine is proposed. Based on the preference of the Human Visual System (HVS), Q-Refine uses the Image Quality Assessment (IQA) metric to guide the refining process for the first time, and modify images of different qualities through three adaptive pipelines. Experimental shows that for mainstream T2I models, Q-Refine can perform effective optimization to AIGIs of different qualities. It can be a general refiner to optimize AIGIs from both fidelity and aesthetic quality levels, thus expanding the application of the T2I generation models.

Q-Refine: A Perceptual Quality Refiner for AI-Generated Image

TL;DR

The paper tackles inconsistent perceptual quality in AI-generated images by introducing Q-Refine, a quality-aware refiner guided by an Image Quality Assessment map. It deploys three adaptive pipelines—Gaussian noise, patch-based inpainting, and global enhancement—to selectively refine low-, medium-, and high-quality regions, respectively. Across multiple AIGI quality datasets and T2I models, Q-Refine demonstrates broad improvements in both fidelity and aesthetic quality, without degrading already high-quality regions. This perceptual, region-specific refinement approach broadens the practical utility of T2I systems by delivering consistent, quality-aware improvements.

Abstract

With the rapid evolution of the Text-to-Image (T2I) model in recent years, their unsatisfactory generation result has become a challenge. However, uniformly refining AI-Generated Images (AIGIs) of different qualities not only limited optimization capabilities for low-quality AIGIs but also brought negative optimization to high-quality AIGIs. To address this issue, a quality-award refiner named Q-Refine is proposed. Based on the preference of the Human Visual System (HVS), Q-Refine uses the Image Quality Assessment (IQA) metric to guide the refining process for the first time, and modify images of different qualities through three adaptive pipelines. Experimental shows that for mainstream T2I models, Q-Refine can perform effective optimization to AIGIs of different qualities. It can be a general refiner to optimize AIGIs from both fidelity and aesthetic quality levels, thus expanding the application of the T2I generation models.
Paper Structure (13 sections, 7 equations, 5 figures, 3 tables)

This paper contains 13 sections, 7 equations, 5 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work and Contributions
  3. Proposed Method
  4. Framework
  5. IQA Module
  6. Stage 1 Pipeline: Gaussian Noise
  7. Stage 2 Pipeline: Mask Inpainting
  8. Stage 3 Pipeline: Global Enhancement
  9. Expriment
  10. Expriment Settings
  11. Expriment Result and Discussion
  12. Ablation Study
  13. Conclusion

Figures (5)

  • Figure 1: The original AIGIs from AGIQA-3Kdatabase:agiqa-3k, optimized by Traditional Refiners and Q-Refine we proposed. As a quality-aware metric, the Q-Refine can add details on the blurred part, to better optimize low-quality regions of (1)(2); improve clarity in medium-quality regions of (3)(4) without changing the whole image; and avoid degrading the high-quality regions of (5)(6).
  • Figure 2: Framework of Q-Refine, including a quality pre-prossess module, and three refining pipelines for low/medium/high quality (LQ/MQ/HQ) regions. The refining mechanisms for each pipeline are inspired by the predicted quality.
  • Figure 3: The refining result by only denoise / add noise + denoise from SDXLgen:XL. Adding noise reduces quality iqa:paq, but it lays the foundation for global optimality before denoising.
  • Figure 4: Using original patch quality map / flattened map to guide the inpainting. (a) suffers from block effects and unexpected artifacts while (b) has a smooth and natural result.
  • Figure 5: Using blind enhancer or prompt-guided enhancer to refine images in different quality groups in AGIQA-3Kdatabase:agiqa-3k. Blind enhancer shows better refining results for LQ groups but causes negative optimization for HQ groups.