Q-Probe: Scaling Image Quality Assessment to High Resolution via Context-Aware Agentic Probing

TL;DR

Q-Probe tackles the difficulty of scaling image quality assessment to high-resolution imagery by marrying context-aware agentic probing with a coarse-to-fine training regime. It introduces Vista-Bench for fine-grained local degradation analysis and employs a three-stage curriculum (perception alignment, hybrid-resolution SFT, and decoupled post-RL) to mitigate cropping-induced biases and misinterpretations of depth-of-field. The approach uses GRPO-based global alignment, a Data Flywheel to generate balanced training trajectories, and decoupled rewards to jointly optimize scoring accuracy and localization precision, achieving state-of-the-art results on Vista-Bench (SRCC ≈ 0.728, PLCC ≈ 0.776) and strong generalization across resolutions. This work provides a scalable, interpretable framework for high-resolution IQA that can enhance MLLM-driven perceptual alignment and quality control in real-world, high-detail imaging applications.

Abstract

Reinforcement Learning (RL) has empowered Multimodal Large Language Models (MLLMs) to achieve superior human preference alignment in Image Quality Assessment (IQA). However, existing RL-based IQA models typically rely on coarse-grained global views, failing to capture subtle local degradations in high-resolution scenarios. While emerging "Thinking with Images" paradigms enable multi-scale visual perception via zoom-in mechanisms, their direct adaptation to IQA induces spurious "cropping-implies-degradation" biases and misinterprets natural depth-of-field as artifacts. To address these challenges, we propose Q-Probe, the first agentic IQA framework designed to scale IQA to high resolution via context-aware probing. First, we construct Vista-Bench, a pioneering benchmark tailored for fine-grained local degradation analysis in high-resolution IQA settings. Furthermore, we propose a three-stage training paradigm that progressively aligns the model with human preferences, while simultaneously eliminating causal bias through a novel context-aware cropping strategy. Extensive experiments demonstrate that Q-Probe achieves state-of-the-art performance in high-resolution settings while maintaining superior efficacy across resolution scales.

Figures (4)

• Figure 1: Challenges in detecting subtle distortions via global perception versus local zooming. (a) Existing MLLMs fail to capture subtle local artifacts. (b) Even when visible via cropping, Semantic Robustness Bias causes models to ignore defects in key semantic areas (e.g., face). (c-d) Naive zooming leads to Logic Collapse, where the model misinterprets natural bokeh as blur (c) or falsely learns that Zooming implies Low Quality (d). (e) Q-Probe mimics human active viewing (Global Perception $\rightarrow$ Local Scrutiny $\rightarrow$ Critical Thinking) to correctly distinguish artifacts from natural effects. (f-g) Data distribution of the high-resolution Vista-Bench and performance comparison showing Q-Probe's superiority.
• Figure 2: This diagram illustrates the construction pipeline of Vista-Bench and the Data Flywheel for SFT. Specifically, we utilize wavelet transforms to decouple structure from texture, selectively injecting artifacts into texture-rich semantic regions, while employing Gemini-2.5 Pro to generate importance-weighted annotations for fine-grained perception probing. To support SFT, we generate traces that interleave global overviews, defect zooming, and context verification (scrutinizing clear regions), thereby preventing the model from associating tool usage solely with defects.
• Figure 3: Overview of the three-stage training framework. Initially, RL Pre-training leverages ranking rewards to align global perception with human preferences. Subsequently, hybrid-resolution SFT enables the model to acquire robust logical reasoning. Finally, the RL Post-training stage fine-tunes the model for precise degradation detection and adaptive tool invocation.
• Figure 4: To monitor the training dynamics, we calculated the average standard deviation of predicted scores across multiple inference runs at various checkpoints. The observed monotonic decrease in variance not only confirms that Q-Probe achieves greater stability throughout Stage-1.