PaperlayerView exact revision on arXiv

Source-linked AI summary

MR-IQA-2: Faithful Image Quality Reflection via Fine-Grained Credit Assignment

Yuan li, Youyuan Lin, Chenhui Chu, Shin'ya Nishida

arXiv:2608.18579v1cs.CVcs.AI

TL;DR

Faithful reasoning in blind image quality assessment remains difficult to verify, especially when reasoning receives credit solely through rating accuracy. MR-IQA-2 uses visual reflection and fine-grained credit assignment to supervise reasoning separately from ratings, retaining competitive human rating alignment while producing more reliable quality reasoning.

  • Problem

    Blind IQA methods can generate plausible explanations without verifying whether identified factors genuinely limit image quality, while rating-based rewards may credit incorrect reasoning.

  • Method

    MR-IQA-2 combines an actor-editor-judge framework with visual reflection and masked token-specific updates that separately supervise reasoning and rating.

  • Results

    MR-IQA-2 retains competitive rating alignment while producing more reliable and actionable quality reasoning.

  • Takeaways & Limitations

    The framework offers a way to evaluate reasoning faithfulness through intervention-generated image pairs rather than rating performance alone.

  • Takeaways & Limitations

    The framework does not yet close the visual-reflection loop by returning edited images to the Actor, and separate modules add computational and parameter redundancy.

Abstract

from arXiv · show

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. Code is available at https://github.com/RobinY99/MR-IQA-2.

1 Introduction

BIQA has shifted from estimating synthetic degradation intensity toward understanding complex real-world images, but MLLM explanations may remain plausible without faithfully identifying quality-limiting factors. MR-IQA-2 addresses this gap by decoupling reasoning and rating supervision and constructing verifiable supervision for faithful reasoning.

  • Motivation: BIQA has moved from degradation-scale estimation toward understanding real-world images with diverse content, capture conditions, aesthetics, and semantic preferences.Synthetic distortion types cannot fully describe these factors, motivating image-centered quality understanding.
  • Problem: MLLMs can generate fluent quality explanations that sound plausible while failing to identify the visual limitations that truly determine image quality.This creates risks of causal reasoning collapse, unverifiable explanations, and deviations from human perception.
  • Problem: Tool-augmented BIQA improves visual inspection and grounding but does not verify whether identified factors genuinely constrain overall image quality.Region-aware observation helps inspect details more reliably, while the faithfulness of proposed quality factors remains unclear.
  • Method and contributions: MR-IQA-2 uses fine-grained credit assignment to decouple reasoning and rating supervision through masked token-specific updates while preserving reasoning-to-rating causality.The framework also constructs verifiable supervision for faithful reasoning through visual reflection and investigates downstream uses including image editing and visual recognition.

2 Related Work

Related work has improved BIQA interpretability through language-based image understanding and task-specific visual evidence, but faithful quality reasoning remains challenging. Existing methods span supervised fine-tuning, reinforcement learning, degradation modeling, and visual tools.

  • MLLM-based BIQA improves interpretability by enabling language-based image understanding and quality explanations, while faithful quality reasoning remains a key challenge.
  • Early methods such as DepictQA and Q-Instruct mainly use supervised fine-tuning to encourage plausible explanation patterns.Their training-data constraints can cause templated responses and limited logical coherence.
  • Later reinforcement-learning frameworks include Q-Insight, VisualQuality-R1, and H-IQA.
  • Before MLLM-based BIQA, degradation modeling and task-specific image processing derived quality cues through contrastive views and distortion-aware representations.Recent methods add region-aware zooming, magnification, and gamma correction for detail inspection and visual evidence.

3 Methods

MR-IQA-2 uses an actor-editor-judge framework to make image-quality reasoning verifiable through targeted edits and independent quality comparisons. Fine-grained credit assignment separately supervises reasoning, rating, and output format while preserving the causal link from reasoning to rating.

  • Actor–Editor–Judge: MR-IQA-2 has an MLLM actor generate quality reasoning and a rating, a frozen diffusion-based editor apply conditioned edits, and a frozen judge assess quality changes.The framework evaluates reasoning through visual intervention rather than rating performance alone.
  • Masked credit assignment: Masked credit assignment updates reasoning tokens with reasoning rewards, rating tokens with rating rewards, and the complete output with the format reward.Each reward is normalized independently within the K samples for an image before constructing its masked advantage.
  • Reasoning reward: Reasoning faithfulness is rewarded by the Judge-score improvement produced when the editor applies low-level quality-attribute interventions while preserving image content.Reasoning that produces a larger quality improvement receives a higher reward.
  • Rating and format rewards: Rating supervision uses relative quality margins against human mean opinion scores, while a separate format reward enforces valid JSON with ordered reasoning and rating fields.The rating reward follows MR-IQA’s relative-margin formulation, and the format contract is denoted F.
  • Optimization: The Actor is optimized with masked Group Relative Policy Optimization, using clipped likelihood ratios and capped token-level mismatch correction.The likelihood ratio is clipped to [1−ϵ_low, 1+ϵ_high], and the correction factor is capped by ω_max.

4 Experiments

Experiments evaluate MR-IQA-2 on in-distribution and out-of-distribution IQA benchmarks, measuring both rating alignment and reasoning faithfulness through editing-based quality gains. Results show competitive ratings, improved and more stable reasoning, and evidence that fine-grained credit assignment separates rating convergence from faithful image-conditioned reasoning.

  • Datasets: MR-IQA-2 is evaluated on KonIQ-10k, SPAQ, LIVE In the Wild, and AI-generated image benchmarks, using KonIQ-10k training and test splits for in-distribution assessment.The KonIQ-10k training split contains 7,046 images, while its test split contains 2,010 images.
  • Rating performance: The frozen-vision variant reaches an average PLCC/SRCC of 0.840/0.816, comparable to Zoom-IQA’s 0.840/0.819 without additional data.MR-IQA-2 uses RL training only, whereas Zoom-IQA uses a two-stage SFT–RL pipeline with additional annotations.
  • Reasoning evaluation: Quality gain from reasoning-guided editing provides a unified proxy for reasoning faithfulness rather than relying on rating performance or limited human inspection.The criterion evaluates image-quality improvement produced by reasoning-guided editing.
  • Cross-editor/judge stability: MR-IQA-2 retains larger quality gains across alternative Editors and Judges, indicating that its reasoning is not specific to a single frozen Editor–Judge pair.The cross-editor/judge analysis replaces FLUX.2 [klein] with Mage-Flow and tests Q-Insight and GPT-5.6 Sol as alternative Judges.
  • Ablation study: Rating-only training achieves an average PLCC/SRCC of 0.836/0.815 but a mean quality gain of −0.260, below the baseline’s −0.136.The result demonstrates that competitive rating alignment alone does not guarantee faithful reasoning.
  • Ablation study: At step 30, the credit mask reaches PLCC/SRCC of 0.745/0.772 versus 0.667/0.617 without it, while quality gain is 0.255 versus 0.387.The credit mask accelerates rating convergence but may temporarily limit reasoning improvement.

5 Discussion

The framework jointly treats evidence and solution as reasoning, but a model may identify an image-quality problem correctly while failing to propose an effective correction. This raises whether training for effective solutions also improves evidence capability.

  • Good Evidence, Bad Solution?: The authors note that correct quality evidence may coexist with an ineffective correction, making it unclear whether solution-oriented exploration improves evidence capability.Training necessarily encourages discovering effective solutions through exploration, but the passage questions whether this process also strengthens evidence.

6 Limitations and Future Work

MR-IQA-2 currently lacks a fully closed-loop visual-reflection process and incurs computational and parameter redundancy from separate Actor, Editor, and Judge models. Future work will feed edited images back to the Actor and investigate smaller models and parameter sharing for greater efficiency.

  • Limitations: The framework does not yet return the edited image to the Actor for further visual reasoning, so visual reflection is not fully closed-loop.Future work will feed the edited image back to the Actor, enabling direct comparison between the original and edited images.
  • Limitations: Using separate Actor, Editor, and Judge models introduces computational and parameter redundancy.
  • Future Work: Future work will explore smaller models and parameter sharing across modules to improve efficiency.

7 Conclusion … A.3 Optimization Algorithms: GRPO or DAPO?

MR-IQA-2 targets visually verifiable image-quality reasoning through an Actor–Editor–Judge framework and fine-grained credit assignment. Supplementary evaluations define dataset protocols, compare lightweight backbones, and show DAPO improves optimization outcomes but GRPO remains the final default because of its lower training cost.

  • 7 Conclusion: MR-IQA-2 uses quality gain to operationalize reasoning faithfulness in an Actor–Editor–Judge framework for visually verifiable BIQA.The framework moves beyond plausible explanations toward visually supported understanding.
  • 7 Conclusion: The Actor’s reasoning guides editing, the Judge evaluates resulting quality changes, and fine-grained credit assignment separates reasoning from rating supervision.The framework links blind assessment to reference-based verification through intervention-generated image pairs.
  • A.1 Dataset Protocol.: Training uses the KonIQ-10k training subset, while final testing covers KonIQ-10k, SPAQ, LIVE-W, AGIQA-3K, KADID-10k, and CSIQ.A fixed set of 200 KonIQ-10k test images monitors early-stage convergence.
  • A.1 Dataset Protocol.: Zoom-IQA results are unavailable, and Tool-IQA results are not directly comparable because of access and training–testing configuration differences.These constraints apply at manuscript completion under the paper’s evaluation protocol.
  • A.1 Dataset Protocol.: The cross Actor–Judge–Editor evaluation uses 5,866 sampled images because full evaluation would require approximately 23,000 images, four editing operations, and three Judge evaluations per image.The subset samples 1,000 images from each test set except CSIQ, whose 866 images are all used.
  • A.2 Candidate Backbones.: Qwen3.5-2B outperforms Qwen3-VL-2B on five of six datasets for PLCC and SRCC, all six for MAE, and uses 0.81× the training time per epoch.The comparison uses matched frozen-vision DAPO settings.
  • A.2 Candidate Backbones.: The 4B model converges faster and produces more diverse outputs, reaching 0.935/0.924/0.228 PLCC/SRCC/MAE by Epoch 3 and a 41.5% U-score.At Epoch 2, T7 reaches 0.929/0.919/0.566 versus 0.915/0.904/0.744 for 2B T6; T7 also produces 83 distinct ratings and 200 unique completions.
  • A.3 Optimization Algorithms: GRPO or DAPO?: DAPO reduces zero-advantage learner groups from 15.75% to 0% and improves Epoch 3 six-dataset average PLCC/SRCC by 0.006/0.008 while reducing MAE by 0.138.Its improvements occur on five, six, and six datasets for PLCC, SRCC, and MAE, respectively.

A.4 Hyper-Parameter Settings. … B.2 Low-Level Visual Feature Patterns

The hyper-parameter analysis identifies settings that balance generalization, rating diversity, and training cost, while visual analyses show that editing magnitude and low-level feature patterns relate systematically to quality improvement. Judge behavior broadly aligns with human preferences, especially for sharpness, without collapsing edited images to fixed feature values.

  • A.4 Hyper-Parameter Settings.: Local groups of six achieve higher six-dataset average PLCC/SRCC than groups of 48, while groups of 48 perform better on in-domain validation alignment.Local-6 obtains 0.835/0.813 versus 0.828/0.809 for six-dataset average PLCC/SRCC; group-48 obtains 0.929/0.921/0.261 versus 0.917/0.906/0.587 for validation PLCC/SRCC/MAE.
  • A.4 Hyper-Parameter Settings.: Active vision training improves authentic-dataset alignment, whereas freezing the vision encoder and aligner performs better on AGIQA-3K and synthetic datasets.Active training reaches 0.952/0.938 on KonIQ and 0.897/0.877 on LIVE-W, while freezing reaches 0.826/0.768 on AGIQA-3K, 0.665/0.676 on KADID-10k, and 0.840/0.801 on CSIQ.
  • A.4 Hyper-Parameter Settings.: E3 U-scores range from 12.0% to 41.5%, corresponding to 24–83 distinct ratings, indicating no observed rating-diversity collapse under the current settings.This contrasts with the 2.04% unique-score ratio reported for Zoom-IQA under a ranking reward.
  • A.4 Hyper-Parameter Settings.: Four repeated policy updates improve validation PLCC/SRCC/MAE but leave six-dataset averages nearly unchanged and more than double training time per epoch.Validation alignment improves from 0.890/0.881/0.691 to 0.910/0.904/0.265, while training time rises from 1.13 to 2.47 hours per epoch.
  • B Visual Quality Feature Analysis: The FLUX.2 [klein] Editor creates original–edited image pairs conditioned on quality reasoning, bridging single-image and reference-based quality assessment.The analysis examines visual-change effects, edited low-level feature distributions, and behavioral preferences.
  • B.1 Visual Change and Quality Gains: Quality gain is negatively correlated with PSNR and MS-SSIM but positively correlated with LPIPS, indicating that larger editing changes generally produce greater quality gains.The reported correlations are ρ = −0.33 to +0.01 for PSNR, ρ = −0.49 to −0.28 for MS-SSIM, and ρ = 0.37 to 0.56 for LPIPS.
  • B.2 Low-Level Visual Feature Patterns: Judge feature-dependent curves broadly follow human-MOS curves across six datasets, with especially similar sharpness and contrast trends, supporting human-like Judge feedback.Sharpness shows the clearest cross-dataset pattern: both human MOS and Judge scores generally increase with sharpness, while other attributes are weaker, nonlinear, or dataset-dependent.
  • B.2 Low-Level Visual Feature Patterns: Edited images retain broad distributions across all six low-level attributes rather than collapsing to fixed values, suggesting the framework does not overfit a single low-level pattern.The passage notes that joint attribute distributions and their relations to human vision and learned perceptual representations require further examination.

B.3 Framework Behavior Preference

The framework adapts its edits to input quality and degradation characteristics, with lower-quality images receiving larger Judge-rated gains and larger visual changes. Judge feedback also shifts reasoning toward perceptual plausibility, while token concentration suggests emerging pattern overfitting; these associations do not establish causality.

  • Source quality, visual change, and quality gain: Lower-MOS images generally receive larger Judge-rated quality gains and tend to have lower original-edited similarity, indicating larger visual changes.The passage links larger visual changes with higher quality gains, while noting that these observed associations do not by themselves establish causality.
  • Preferences over feature changes: Changes in sharpness show the strongest and most consistent relationship with quality improvement, while preferred corrections for entropy and saturation depend on the degradation domain.Authentic and synthetic datasets exhibit different response patterns for entropy and saturation.
  • Behavioral adaptation and concentration: After training, reasoning increasingly uses terms such as natural and realistic, indicating a learned preference for perceptually plausible corrections.The framework adjusts its reasoning vocabulary in response to Judge feedback.
  • Behavioral adaptation and concentration: The token restore accounts for 20.6% of displayed frequency, while the top-10 token share reaches 44.1%, suggesting concentration on a small set of solution patterns.The passage characterizes this concentration as an emerging tendency to overfit a small set of solution patterns.

C Discussion on the Editor Proxy

The Editor proxy does not disentangle instruction-following capability from semantic preservation, creating attribution risks that can weaken reasoning-faithfulness supervision. Future work should independently evaluate instruction following, semantic consistency, and perceptual quality.

  • Limitations: The Editor proxy conflates instruction-following capability with semantic preservation, limiting attribution of observed quality changes.Quality may degrade because of Editor limitations rather than incorrect Actor instructions, while semantic inconsistency can weaken quality change as a proxy for reasoning faithfulness.
  • Future Work: Future work should use independent human or human-like evaluation to assess instruction following, semantic consistency, and perceptual quality separately.This separation is intended to provide more reliable feedback and a more faithful framework.

D Joint Rating, Reasoning, and Diversity Analysis · E Detailed Credit-Mask Audit · E.1 Audit Scope and Protocol

The analysis jointly evaluates rating alignment, functional reasoning, and diversity, showing that credit masking trades a small amount of rating strength for more useful and diverse reasoning. A controlled audit further verifies distinct reward routing, complete execution, and eligibility-related edge cases.

  • D Joint Rating, Reasoning, and Diversity Analysis: Evaluation averages ratings across six datasets and 28,270 test images, while reasoning uses Judge-score gain on 5,843 common image keys and diversity covers 23,599 successful inputs.The reasoning audit uses a fixed FLUX Editor and E5 Judge and excludes 23 length-limited Mask generations.
  • D Joint Rating, Reasoning, and Diversity Analysis: Without Mask has the lowest average MAE in the controlled E1 comparison, whereas Mask + KL recovers most of Mask’s correlation gap and gives the strongest LIVE-W result.At E5, Mask has the best average PLCC and performs better on most authentic and AI-generated datasets, while Without Mask is stronger on KADID-10k and slightly higher in CSIQ SRCC.
  • D Joint Rating, Reasoning, and Diversity Analysis: The rating-only Actor achieves strong rating alignment but reduces quality on average; Without Mask has strongest E1 rating correlation, while Mask produces the largest quality gain.Mask + KL recovers most of the rating gap while retaining a positive mean gain, separating rating accuracy from functional reasoning usefulness.
  • D Joint Rating, Reasoning, and Diversity Analysis: Without Mask collapses evidence to three distinct strings, with one accounting for 99.992% of matched outputs, whereas Mask increases unique evidence strings to 291 and yields the highest solution diversity.Evidence and solution diversity must therefore be evaluated separately because one field can collapse while the other remains varied.
  • D Joint Rating, Reasoning, and Diversity Analysis: At E5, Without Mask retains 74.719% unique evidence but has 0.004% solution uniqueness and a 100% solution top-1 share, demonstrating a field-wise diversity mismatch.Exact uniqueness is only a lexical proxy and does not establish semantic diversity or faithfulness.
  • D Joint Rating, Reasoning, and Diversity Analysis: Without Mask E5 reaches +1.431 raw gain but maps all 28,270 inputs to one normalized solution, whereas Mask E5 achieves 0.838/0.812 average PLCC/SRCC and +1.071 mean quality gain.Mask also produces 19,618 unique evidence strings and 23,457 normalized solutions across 28,044 successful full-audit cases.
  • E Detailed Credit-Mask Audit: The audit compares two complete five-epoch runs: Mask routes rewards to supervised outputs with output-specific KL, while Without Mask broadcasts rewards globally; both use KL coefficient 0.02.Each run has 1,455 steps and 209,520 trajectories, for 419,040 total, with six samples per image and 160token completions.
  • E.1 Audit Scope and Protocol: No audit step or record is missing, duplicated, invalid, or non-finite; at Without Mask steps 711 and 713, all samples are Actor-ineligible, so absent Judge changes reflect eligibility.Both runs use frozen vision encoder and aligner, and the Editor consumes only the solution without an explicit semantic guardrail.

E.2 Training Dynamics … E.9 Causal Scope

Across checkpoints, both configurations improve rating alignment, but Without Mask develops semantic and lexical solution collapse despite strong rewards and ratings. Mask changes more gradually, while the comparison cannot isolate whether credit routing or KL scope causes the failure.

  • E.2 Training Dynamics: Mask’s reasoning reward rises from 0.207 to 0.307 and Judge gain from 0.586 to 0.840, while rating rewards converge by E5.Without Mask shows larger reasoning reward and Judge gain, but shorter outputs and solution collapse; its final-window slopes indicate a plateau and slight decline.
  • E.3 Checkpoint Validation: Without Mask maintains a Judge gain near 1.18 while normalized-solution uniqueness falls from 91.88% at E1 to about 0.5% at E2.The semantic dwelling-template rate reaches 100% at E1, preceding near-exact lexical collapse.
  • E.4 Cross-Dataset Outcomes: At E5, Mask has similar six-dataset PLCC/SRCC averages to Without Mask (0.838/0.812 versus 0.834/0.812), while Without Mask has larger pooled quality gain (1.431 versus 1.071).All 28,270 Without Mask outputs share one normalized solution; Mask retains 23,457 among 28,044 successful outputs.
  • E.5 Collapse Milestones: Under Without Mask, the semantic dwelling family exceeds 90% at step 265, whereas one normalized sentence does not exceed 90% until step 551.Semantic detectors therefore identify shared concepts before normalized near-duplicate matching identifies one sentence.
  • E.6 Output-Length Evolution: Mean training completion decreases from 89.39 to 85.26 tokens under Without Mask but rises from 100.57 to 109.00 under Mask.Validation solution lengths change from 32.56 to 18.00 and from 37.85 to 50.94, respectively; the contraction is consistent with convergence to one reusable instruction.
  • E.7 Rating and Diversity Diagnostics: Without Mask retains KonIQ PLCC/SRCC of 0.932/0.914 despite solution collapse, because all 2,010 solutions are identical after normalization.Rating correlation and full-output uniqueness therefore miss collapse, requiring inspection of supervised outputs, image–solution relevance, and semantic preservation.
  • E.8 Representative Validation Outputs: Different evidence and ratings can yield the same normalized solution: Sample 1 receives 0.865 reward when a fixed intervention raises Judge score from 2.37 to 4.37.This measures edited-image quality without establishing source-image faithfulness.
  • E.9 Causal Scope: The comparison establishes a reproducible Without Mask/global KL failure mode but cannot identify the responsible change.Causal attribution requires a 2×2 grid crossing Without Mask versus Mask with global versus component KL, using multiple seeds.
Loading 2608.18579v1…