Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

iDETEX: Empowering MLLMs for Intelligent DETailed EXplainable IQA

Zhaoran Zhao, Xinli Yue, Jianhui Sun, Yuhao Xie, Tao Shao, Liangchao Yao, Fan Xia, Yuetang Deng

TL;DR

iDETEX introduces a unified multimodal large language model for detailed, explainable IQA, jointly addressing quality grounding, perception, and description. It leverages three task-specific offline augmentations plus a data mixing strategy and online high-resolution enhancements, with a foundation-model replacement to InternVL3, achieving state-of-the-art results on ViDA-UGC and rank-1 in the MIPI 2025 Detailed IQA Challenge. Key contributions include targeted augmentation modules, a task-aware mixing scheme, and online strategies that improve localization, perceptual sensitivity, and causal description of image quality. The results demonstrate robust, interpretable IQA performance with potential for scalable transfer learning to low-data regimes in diverse distortion settings.

Abstract

Image Quality Assessment (IQA) has progressed from scalar quality prediction to more interpretable, human-aligned evaluation paradigms. In this work, we address the emerging challenge of detailed and explainable IQA by proposing iDETEX-a unified multimodal large language model (MLLM) capable of simultaneously performing three key tasks: quality grounding, perception, and description. To facilitate efficient and generalizable training across these heterogeneous subtasks, we design a suite of task-specific offline augmentation modules and a data mixing strategy. These are further complemented by online enhancement strategies to fully exploit multi-sourced supervision. We validate our approach on the large-scale ViDA-UGC benchmark, where iDETEX achieves state-of-the-art performance across all subtasks. Our model ranks first in the ICCV MIPI 2025 Detailed Image Quality Assessment Challenge, demonstrating its effectiveness and robustness in delivering accurate and interpretable quality assessments.

iDETEX: Empowering MLLMs for Intelligent DETailed EXplainable IQA

TL;DR

iDETEX introduces a unified multimodal large language model for detailed, explainable IQA, jointly addressing quality grounding, perception, and description. It leverages three task-specific offline augmentations plus a data mixing strategy and online high-resolution enhancements, with a foundation-model replacement to InternVL3, achieving state-of-the-art results on ViDA-UGC and rank-1 in the MIPI 2025 Detailed IQA Challenge. Key contributions include targeted augmentation modules, a task-aware mixing scheme, and online strategies that improve localization, perceptual sensitivity, and causal description of image quality. The results demonstrate robust, interpretable IQA performance with potential for scalable transfer learning to low-data regimes in diverse distortion settings.

Abstract

Image Quality Assessment (IQA) has progressed from scalar quality prediction to more interpretable, human-aligned evaluation paradigms. In this work, we address the emerging challenge of detailed and explainable IQA by proposing iDETEX-a unified multimodal large language model (MLLM) capable of simultaneously performing three key tasks: quality grounding, perception, and description. To facilitate efficient and generalizable training across these heterogeneous subtasks, we design a suite of task-specific offline augmentation modules and a data mixing strategy. These are further complemented by online enhancement strategies to fully exploit multi-sourced supervision. We validate our approach on the large-scale ViDA-UGC benchmark, where iDETEX achieves state-of-the-art performance across all subtasks. Our model ranks first in the ICCV MIPI 2025 Detailed Image Quality Assessment Challenge, demonstrating its effectiveness and robustness in delivering accurate and interpretable quality assessments.
Paper Structure (32 sections, 5 equations, 2 figures, 11 tables)

This paper contains 32 sections, 5 equations, 2 figures, 11 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. Spatial Perturbation Augmentation for Grounding
  5. Random Cropping
  6. Horizontal Flipping
  7. Adjustment of Bounding Box Coordinates
  8. Query-Style Aligned Augmentation for Perception
  9. Score-Aware Augmentation for Description
  10. Score-Driven Inference Simplification
  11. Granularity-Aware Label Refinement
  12. Original Method:
  13. Enhanced Method:
  14. Task-Aware Augmented Data Mixing
  15. Global Augmentation Strategies
  16. ...and 17 more sections

Figures (2)

  • Figure 1: Illustration of the input and output formats for the three core tasks for explainable IQA in ViDA-UGC DYEvalab2025ViDAMIPI, using representative examples. From left to right: Image Quality Grounding, Image Quality Perception, and Image Quality Description.
  • Figure 2: Overview of our training paradigm for iDETEX. Three task-specific augmentation strategies are proposed—Spatial Perturbation Augmentation, Query-Style Aligned Augmentation, and Score-Aware Augmentation—designed for the grounding, perception, and description tasks, respectively. Based on these, we introduce a Task-Aware Augmented Data Mixing Strategy that enables multi-source data to effectively participate in MLLM fine-tuning. These offline augmentations, together with an online High-Resolution Input Enhancement strategy, form a unified fine-tuning pipeline. The augmented data is used to fine-tune a replaced base MLLM, resulting in iDETEX, a model capable of delivering accurate, robust, and interpretable quality assessments across diverse distortion scenarios.