文章背景与核心概要
近年来,多模态大语言模型(MLLMs)在图像质量评估(IQA)领域展现出了巨大潜力,它们通过将质量评分与底层推理过程进行对齐,提升了评估的可解释性。然而,传统方法通常仅通过人类提供的最终评分来监督推理过程,而未能验证这些推理是否真实反映了图像的实际质量。因此,评分准确并不能保证推理的真实性,且共享的奖励机制还可能在碰巧得到正确评分时,反而强化了不真实的推理逻辑。
为了克服这一局限性,本文介绍了 MR-IQA-2,这是一个创新性的“演员-编辑-裁判”(actor-editor-judge)框架。该框架实现了两大核心突破:一是解耦了推理与评分的信用分配(credit assignment),二是利用视觉反射(visual reflection)和细粒度更新为真实推理提供了可验证的监督。通过这一机制,MR-IQA-2 不仅在各项 IQA 基准测试中实现了与人类评分的高度对齐,还赋予了模型超越单纯评分的、更丰富且更真实的视觉理解能力。
MR-IQA-2:通过细粒度信用分配实现真实的图像质量反映
摘要
Multimodal large language models (MLLMs) have demonstrated significant potential for image quality assessment (IQA) by aligning quality ratings with underlying reasoning. However, traditional approaches often supervise reasoning solely through human-provided ratings without verifying whether the reasoning faithfully reflects actual image quality. Consequently, rating accuracy does not guarantee faithful reasoning, and shared reward mechanisms can reinforce unfaithful logic when correct ratings happen by chance.
To resolve this limitation, MR-IQA-2 introduces an actor-editor-judge framework that: 1. Decouples credit assignment for reasoning and rating. 2. Provides verifiable supervision for faithful reasoning via visual reflection and fine-grained updates.
多模态大语言模型(MLLMs)通过将质量评分与其底层的推理过程相契合,在图像质量评估(IQA)领域展现出巨大的潜力。然而,大多数方法仅通过人类提供的评分来监督推理,极少检查推理过程是否真实反映了图像的质量。单纯的评分准确并不能保证推理的真实性;共享的奖励机制还掩盖了监督来源,当碰巧产生正确评分时,甚至可能会强化不真实的推理逻辑。为了提高盲图像质量评估(blind IQA)的真实性与可靠性,我们的目标是:(1)解耦推理与评分的信用分配,以及(2)为真实推理提供可验证的监督。为此,我们引入了 MR-IQA-2,这是一个将“推理-编辑-反射”操作化的演员-编辑-裁判(actor-editor-judge)框架。演员为输入图像生成质量推理,编辑器根据识别出的质量因素对图像进行修改。冻结的裁判对比原始图像与编辑后的图像,并为演员的推理提供反射式监督。MR-IQA-2 进一步利用细粒度信用分配来解耦推理与评分的监督。裁判的反馈监督推理,而人类评分则监督预测评分。掩码特定的标记更新能够区分这些信号,同时保留从推理到评分的因果关系。在各项 IQA 基准测试中,MR-IQA-2 实现了极具竞争力的与人类评分的对齐效果。视觉反射还带来了超越评分的、更丰富且更真实的视觉理解,这有望为图像质量优化及相关的下游任务提供支持。
论文元数据
- arXiv ID: arXiv:2608.18579 [cs.CV]
- 研究领域: 计算机视觉与模式识别 (
cs.CV); 人工智能 (cs.AI) - 提交日期: 2026年8月19日
- 作者: Yuan Li, Youyuan Lin, Chenhui Chu, Shin'ya Nishida
- 官方代码库: GitHub - MR-IQA-2
摘要
Multimodal large language models (MLLMs) have shown strong potential for image quality assessment (IQA) by improving consistency between quality ratings and their underlying reasoning. However, most approaches supervise reasoning through human-provided ratings and rarely examine whether it faithfully reflects image quality. Rating accuracy alone does not ensure faithful reasoning; a shared reward also obscures supervision sources and may reinforce unfaithful reasoning when a correct rating occurs by chance. To improve the faithfulness and reliability of blind IQA, we aim to (1) decouple credit assignment for reasoning and rating and (2) provide verifiable supervision for faithful reasoning. We introduce MR-IQA-2, an actor-editor-judge framework that operationalizes reasoning-editing-reflection. The actor generates quality reasoning for an input image, and the editor revises the image according to the identified quality factors. A frozen judge compares the original and edited images and provides reflective supervision for the actor's reasoning. MR-IQA-2 further uses fine-grained credit assignment to decouple reasoning and rating supervision. Judge feedback supervises reasoning, whereas human ratings supervise the predicted rating. Masked token-specific updates distinguish these signals while preserving the causal relation from reasoning to rating. Across IQA benchmarks, MR-IQA-2 achieves competitive rating alignment with humans. Visual reflection also enables richer and more faithful visual understanding beyond rating, which may inform image-quality optimization and related downstream tasks.
多模态大语言模型(MLLMs)通过提高质量评分与其底层推理之间的一致性,在图像质量评估(IQA)中表现出强劲的潜力。然而,大多数方法通过人工提供的评分来监督推理,很少检验推理是否真实地反映了图像质量。单纯的评分准确并不能确保推理的真实性;共享奖励还会模糊监督来源,并在巧合产生正确评分时强化不真实的推理。为了提高盲图像质量评估的真实性和可靠性,我们旨在:(1)解耦推理和评分的信用分配,以及(2)为真实推理提供可验证的监督。我们引入了 MR-IQA-2,这是一个将推理-编辑-反射操作化的演员-编辑-裁判框架。演员为输入图像生成质量推理,编辑器根据识别出的质量因素修改图像。一个冻结的裁判比较原始图像和编辑后的图像,为演员的推理提供反射式监督。MR-IQA-2 进一步利用细粒度信用分配来解耦推理和评分监督。裁判反馈监督推理,而人类评分监督预测评分。掩码特定的标记更新区分了这些信号,同时保持了从推理到评分的因果关系。在各个 IQA 基准测试中,MR-IQA-2 实现了与人类具有竞争力的评分对齐。视觉反射还实现了超越评分的更丰富、更真实的视觉理解,这可以为图像质量优化及相关下游任务提供参考。
核心框架:演员-编辑-裁判(Actor-Editor-Judge)
- 演员(The Actor): 为输入图像生成结构化的质量推理和预测评分。
- 编辑器(The Editor): 根据演员识别出的具体质量因素修改图像。
- 裁判(The Judge,冻结): 对比原始图像和编辑后的图像,直接向演员的推理过程提供反射式监督。
全文与资源
