Zoom-IQA: Image Quality Assessment with Reliable Region-Aware Reasoning¶

Conference: ECCV 2026
Paper: ECCV page · Project page
Area: Multimodal VLM / Vision-Language Model Reasoning
Keywords: image quality assessment, region-aware reasoning, visual grounding, reinforcement learning, KL-Coverage

TL;DR¶

Zoom-IQA equips a no-reference image quality assessment model with selective cropping and re-inspection, combining GR-IQA supervision with stabilized reinforcement learning to improve explanation quality under model-based evaluation while retaining competitive score correlations, rather than claiming the best scores on every benchmark.

Background & Motivation¶

Image quality assessment supports not only photo scoring but also perceptual evaluation of restoration and generation models. Q-Align and DeQA-Score map visual content to scores effectively, but offer limited explanations for deductions; the DepictQA series describes distortions without making precise continuous scoring its primary objective. Q-Insight and VisualQuality-R1 use reinforcement learning (RL) to make a single vision-language model (VLM) produce both rationales and scores, yet their reasoning remains largely a single pass of text generation.

A comprehensive explanation does not necessarily mean the image was inspected properly. A low score may reflect subject motion blur, background overexposure, or missing local texture, and viewing only a resized global image can conflate these phenomena. Human assessors zoom into uncertain regions before deciding how much a defect affects overall quality. Teaching a model to do so is not a straightforward extension of visual question answering: an overall quality score rarely corresponds to a unique object box, and datasets do not record where assessors looked or why they revised their scores.

The paper therefore addresses both data and policy learning. It constructs trajectories containing regional evidence and action decisions, first teaches valid crop execution, and then uses scoring feedback to learn when further inspection is worthwhile. Core idea: turn quality assessment into an interactive process that can acquire additional visual evidence, rather than extending textual reasoning over an unchanged visual input, while using targeted regularization and sampling schedules to prevent RL from collapsing onto a few fixed scores.

Method¶

Overall Architecture¶

Built on Qwen2.5-VL-7b, the model takes an image and a rating question and outputs an explanation, a quality score, and an action. Training first applies supervised fine-tuning (SFT) to approximately 7,000 GR-IQA trajectories, followed by Group Relative Policy Optimization (GRPO) to learn region selection and stopping. The latter uses score labels without requiring additional human-annotated regional trajectories.

At inference time, the model first assesses overall quality and explains what remains uncertain. It chooses final if the evidence is sufficient; otherwise, it emits crop and a bounding box. A tool returns the cropped and zoomed region, allowing the model to check its earlier judgment against local evidence. KL-Coverage, rewards, and progressive re-sampling constrain training; they are not additional visual modules called sequentially during deployment.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    D["GR-IQA Dual Filtering<br/>Generate and filter training trajectories"] --> S["SFT cold start"]
    S -.Policy initialization.-> P["Self-Guided Regional Re-inspection<br/>Global assessment and action selection"]
    I["Image and rating question"] --> P
    P -->|crop| C["Crop-and-zoom tool"]
    C -->|Local visual evidence| P
    P -->|final| O["Explanation and final score"]
    K["Targeted KL-Coverage Regularization"] -.Training-only constraint.-> P
    R["Joint Rewards and Progressive Re-sampling"] -.Training-only optimization and scheduling.-> P

Solid arrows indicate data or action flow, while dashed arrows denote training-time initialization and supervision. Filtering supports the cold start, self-guided regional re-inspection remains active at deployment, and the last two designs jointly shape GRPO rather than forming a sequential pipeline.

Key Designs¶

1. GR-IQA Dual Filtering: test both whether answers depend on the image and whether rationales agree with it

The authors use Gemini-2.5-pro to generate structured trajectories from KonIQ images. The <think> block contains an overall quality summary, directions for improvement, issues to avoid, and a decision rationale; <answer> specifies the action, score, and a bounding box when needed. The generator must explain whether an image is an easy case that permits an immediate conclusion or a hard case requiring a crop. Supervision therefore covers not only defect descriptions but also reasons for observing further or stopping.

Visual Reliance Filtering (VRF) fixes a partial rationale and compares continuations under two conditions: image plus prefix, and prefix alone. It uses signals such as score changes, localization differences, and output uncertainty to reject samples whose answers barely change after removing the image. Importantly, it does not perturb the original image: in IQA, a perturbation directly changes the quality being assessed and is not an innocuous control. The main paper does not specify complete thresholds for these differences, so it does not disclose every rule needed to reproduce this step.

Hint-Augmented Consistency Filtering (HACF) then checks intermediate rationales. The main text names Qwen-2.5-32b as the rater and states that it combines the rationale, image, and low-level hints such as brightness, sharpness, and color information to make a binary acceptance decision. Its role differs from VRF: an image-dependent answer can still contain unsupported descriptions, so the text also needs sample-level filtering. The main paper provides limited detail on how this rater receives images; this note preserves the reported name rather than silently substituting a different multimodal model.

The approximately 7,000 retained trajectories train the base model with autoregressive cross-entropy, teaching output formatting, localization expressions, and crop actions. This is model-assisted quality-rationale data, not a record of human eye movements or actual human decision processes.

2. Self-Guided Regional Re-inspection: turn uncertainty into an additional observation rather than a longer description

SFT teaches region expression and tool use but does not ensure that the model knows where to inspect. Second-stage GRPO therefore samples alternative trajectories and explores useful observation policies through final score accuracy and ranking performance. The action space includes both further cropping and termination. This differs from unconditional center cropping or using a segmentation model to extract fixed regions whose local scores are then averaged.

Crucially, a crop should address an unresolved quality question. An image may contain sharp pavement alongside blurred people; the model needs to establish how much subject detail has been lost, rather than lowering the overall score merely because people are semantically salient. Once the crop returns additional local visual input, the model can confirm its previous judgment or revise the score. The tool itself does not supply a final quality label.

Here, “uncertainty awareness” is expressed mainly through generated rationales and actions. The main text does not introduce a calibrated uncertainty probability or establish that each crop action maximizes expected information gain. The evidence therefore supports useful learned inspection behavior, not an optimal active-perception policy.

3. Targeted KL-Coverage Regularization: prevent a few numerical tokens from rapidly dominating RL updates

The authors observe score collapse in reasoning-based IQA. On the KonIQ test set, with scores rounded to two decimal places, VisualQuality-R1 has a unique-score ratio of only 2.04%, compared with 71.34% for ground-truth mean opinion scores (MOS). This indicates concentration on a limited set of values; it is not a quantitative report of the score diversity achieved by Zoom-IQA itself.

KL-Coverage does not treat all generated tokens equally. For each token, it computes the deviation of its log-probability from the batch mean and the deviation of its advantage from the batch mean, then multiplies the two to obtain a covariance signal. It ranks numerical tokens within <answer> by this signal, selects only a small top fraction—for example, a proportion parameter of 0.02—and applies an old-to-current-policy KL divergence constraint at those positions, averaged over the selected positions. Intuitively, this targets updates that further reinforce already probable tokens carrying high advantage.

The motivation emphasizes score tokens, but the textual definition of the candidate set includes all numerical tokens inside the answer tags. Since the answer also contains coordinates, one cannot assume that the implementation covers only the rating field. Equations (1) and (2) have missing symbols in the cached extraction. This note therefore explains the selection mechanism and KL direction from the surrounding prose rather than presenting a reconstructed equation as the authors’ exact formula.

4. Joint Rewards and Progressive Re-sampling: constrain policy learning through absolute scores, relative ordering, and tail examples

The format reward checks whether the rationale and answer satisfy the required structure, assigning 1.0 when all requirements are met and 0 otherwise. The score reward decays in Gaussian form with the distance between the predicted score and MOS, with a scale parameter controlling error sensitivity. The rank reward compares image pairs within a batch: ground-truth preferences follow MOS ordering, while predicted preferences are obtained through the Thurstone model from the means and variances of each image’s rating distribution. Reward then reflects agreement between predicted and ground-truth preferences. This is not simply a binary bonus for correct ordering; it uses preference probabilities informed by rating distributions.

Score reward constrains absolute values, whereas rank reward constrains relative order, and the two are not interchangeable. In the experiments, rank-only reward favors some synthetic-distortion datasets, while score-only reward favors real photographs. The main text specifies weights of 1 and 2 for the score and rank terms, respectively. However, operators in cached Equations (4)–(6) are incompletely extracted, so this note does not reconstruct the full reward formula or treat an unverified additive or gating relationship as a confirmed statement from the paper.

Training scores also have a long-tailed distribution, with too few very high- and very low-quality examples. Progressive re-sampling first trains on the original distribution, then gradually increases the sampling frequency of scarce score intervals in later stages. It changes how often training examples appear, rather than artificially stretching scores after inference. The main text does not provide all sampling ratios and interval settings needed to reproduce the complete schedule.

A Worked Example¶

Figure 1 shows a photograph containing a rickshaw and pedestrians. In the first turn, the model identifies motion blur as the main issue but wants to verify the loss of facial and clothing detail. It outputs bounding box [0.29, 0.24, 0.88, 0.73], an initial score of 3.56, and the crop action.

After the tool returns the crop, the second turn confirms that motion blur substantially damages facial and clothing detail. The model terminates with final and a score of 3.47, against a ground-truth MOS of 3.51. The example illustrates how a local observation can support score revision, but one successful case cannot establish the causal faithfulness of every rationale or imply a fixed two-turn limit.

Loss & Training¶

SFT uses standard autoregressive cross-entropy over the complete structured response, with Qwen2.5-VL-7b as the base model. The main paper reports batch size 2, 4 gradient accumulation steps, learning rate 2.5×10⁻⁶, and warm-up ratio 0.3.

GRPO continues from the SFT model, using reported batch size 1, 2 gradient accumulation steps, learning rate 1×10⁻⁶, KL penalty coefficient 0.04, and 8 generated responses per group. That KL coefficient is a reported implementation setting; it does not establish an otherwise unspecified targeted-regularizer weight or maximum number of crop turns.

Key Experimental Results¶

Main Results¶

The trainable methods in Table 1 of the paper are all trained on KonIQ, and the authors retrain VisualQuality-R1 using its official code. The other six test sets cover real photographs, synthetic distortions, and AI-generated images. Each cell below reports PLCC / SRCC, both higher-is-better, measuring linear and rank correlation, respectively.

Dataset	DeQA-Score	Q-Insight	VisualQuality-R1	Zoom-IQA
KonIQ	0.953 / 0.941	0.918 / 0.895	0.910 / 0.896	0.938 / 0.922
SPAQ	0.895 / 0.896	0.903 / 0.903	0.889 / 0.892	0.902 / 0.900
KADID	0.694 / 0.687	0.702 / 0.702	0.703 / 0.712	0.701 / 0.700
PIPAL	0.472 / 0.478	0.458 / 0.435	0.451 / 0.441	0.468 / 0.465
LIVE-Wild	0.892 / 0.879	0.870 / 0.839	0.856 / 0.827	0.887 / 0.870
AGIQA	0.809 / 0.729	0.816 / 0.766	0.817 / 0.760	0.816 / 0.765
CSIQ	0.787 / 0.744	0.685 / 0.640	0.768 / 0.707	0.797 / 0.754

Zoom-IQA outperforms the two reasoning-based baselines on KonIQ, PIPAL, LIVE-Wild, and CSIQ, but does not lead across all metrics on SPAQ, KADID, or AGIQA. DeQA-Score still achieves higher correlations on both metrics for KonIQ, PIPAL, and LIVE-Wild. The benefits should therefore be assessed alongside explanation capability, not described as a uniform new scoring state of the art.

Table 2 of the paper uses Gemini-2.5-Flash and GPT-5-mini as model judges, rating descriptions on a 1–9 scale. The table below retains the complete four-dimensional Gemini-2.5-Flash results. Judges receive the image, rationale, and low-level indicators. Detailed scoring definitions are in an appendix not supplied here; “Confidence” in this table should not be interpreted as statistical calibration.

Dataset	Method	Accuracy	Reasonableness	Completeness	Confidence
KonIQ	DepictQA	5.40	5.49	5.51	7.96
KonIQ	VisualQuality-R1	7.29	7.60	7.29	7.57
KonIQ	Q-Insight	7.17	7.44	7.08	6.93
KonIQ	Zoom-IQA	8.72	8.80	8.30	8.60
SPAQ	DepictQA	6.04	6.39	6.14	7.86
SPAQ	VisualQuality-R1	8.32	8.35	7.70	7.55
SPAQ	Q-Insight	7.84	8.02	7.51	6.98
SPAQ	Zoom-IQA	8.63	8.69	8.47	8.63

GPT-5-mini gives the same overall ordering: for example, KonIQ accuracy is 6.93 for Zoom-IQA, 6.10 for VisualQuality-R1, and 5.32 for Q-Insight. The two judges assign different absolute scores, underscoring that these are results under specific judging protocols rather than direct measurements of human-verified truthfulness.

Ablation Study¶

The following table selects PLCC results from the paper’s Table 4. Its last three rows compare joint score-and-rank rewards, the addition of targeted regularization, and then progressive re-sampling. Four representative datasets are retained, without conflating SFT baselines from different settings.

Config	KonIQ	KADID	PIPAL	CSIQ
SFT baseline	0.836	0.632	0.431	0.688
Rank reward + KL-Coverage	0.906	0.709	0.450	0.772
Score reward + KL-Coverage	0.928	0.677	0.379	0.715
Joint rewards without KL-Coverage	0.908	0.683	0.455	0.759
Joint rewards + KL-Coverage	0.932	0.665	0.458	0.791
Add progressive re-sampling: full model	0.938	0.701	0.468	0.797

Both single-reward rows retain KL-Coverage, as checked in the original table, and all rows start from SFT. Adding the regularizer raises KonIQ from 0.908 to 0.932 but lowers KADID from 0.683 to 0.665: benefits on most datasets should not be rewritten as universal improvements. Progressive re-sampling subsequently improves both correlations across all seven datasets in Table 4.

Key Findings¶

Region selection is not another name for center cropping. In Table 5, center cropping achieves KonIQ PLCC / SRCC of 0.898 / 0.876, versus 0.938 / 0.922 for the full policy. With the rest of the pipeline held unchanged, region selection contributes measurable value.
Single-stage imitation does not replace the full training process. Table 5 reports 0.834 / 0.812 for Single-Stage SFT, below the full model. This baseline lacks both the multi-stage zoom pipeline and RL, so it is not a pure RL ablation changing only one training objective. It also differs from the Table 4 SFT baseline of 0.836 / 0.806.
Downstream usefulness comes with a perception–fidelity trade-off. For SUPIR / DIV2K, Zoom-IQA guidance yields MANIQA 0.5455 and CLIPIQA 0.6773, but PSNR is 21.31, below Q-Insight’s 22.44. Better perceptual metrics do not imply improved fidelity.
Transfer across restoration frameworks does not mean winning every metric. On DreamClear / RealLQ250, it achieves NIQE 3.914, MUSIQ 66.97, and CLIPIQA 0.6946. MANIQA is 0.4365, slightly below LLaVA-13b’s 0.4369. In the paper’s Table 3, NIQE belongs to the DreamClear setting, not the SUPIR setting.

Highlights & Insights¶

Image-removal controls suit evaluation tasks whose inputs cannot be freely perturbed. VRF checks whether a rationale can continue without the image rather than damaging the image under assessment. This better respects the measurement target of IQA, although it remains a proxy test of visual dependence.
Stabilization exploits the structure of task outputs. KL-Coverage focuses on numerical answer positions rather than constraining all text equally. The transferable insight is that policy collapse in continuous-scoring tasks can first appear as excessive reinforcement of a small set of numerical tokens.
An explanation can become actionable guidance for another model. Directions for improvement and issues to avoid naturally support restoration prompts, and the paper tests this across frameworks. Downstream performance supports practical usefulness, however, not the truthfulness of every statement.

Limitations & Future Work¶

Reasoning reliability is still evaluated mainly by models. Two judges provide cross-checking, but the generation teacher and one judge belong to the Gemini family, allowing potentially correlated preferences. Independent human assessment, regional evidence annotations, and counterfactual tests would strengthen validation.
Inspection costs are insufficiently characterized. The supplied main text does not systematically report average crop count, maximum interaction budget, latency, or token costs. Benefits from re-inspection should be weighed against additional visual encoding and text generation.
Data and implementation details remain incomplete. GR-IQA is generated from KonIQ images, and the main paper does not fully specify filtering thresholds, the rater’s image interface, or progressive sampling schedules. These descriptions alone do not permit complete training reproduction.
Useful localization is not necessarily causal evidence for a score. The authors additionally analyze crops using a saliency model, but salient regions need not contain the distortions most important to quality judgment. Controlled restoration or masking of selected regions could test whether score changes match the rationale’s claims.
Restoration results do not establish reduced hallucination. Higher perceptual metrics alongside lower PSNR indicate a metric trade-off. Determining whether guidance introduces invented texture requires a separate content-fidelity evaluation.

Compared with Q-Insight / VisualQuality-R1: All use RL to combine scoring and explanation. Zoom-IQA additionally acquires local visual evidence during reasoning and constrains training collapse rather than merely extending textual rationales.
Compared with DeQA-Score / Q-Align: These methods primarily target visual scoring and remain stronger on some correlation metrics. Zoom-IQA is better motivated when an application also needs inspection locations and restoration advice, not as a universal replacement for score-only models.
Compared with DOG-IQA and regional reasoning in visual question answering: DOG-IQA uses preprocessing and regional aggregation to imitate assessment, while interactive regional VQA often exploits localizable semantic evidence. Zoom-IQA addresses holistic quality judgments without a unique answer region, contributing task-specific training data and interaction policies.
Research direction: Evaluate whether new visual evidence causes an appropriate change in quality judgment before and after cropping. Under a fixed tool budget, compare active, random, and central region selection, including uninformative crop controls. This comes closer to measuring the value of observation than judging whether an explanation sounds reasonable.

Rating¶

Novelty: 4/5. The combination of learned regional re-inspection and targeted anti-collapse training is well adapted to IQA, although regional tool-based reasoning has precedents in adjacent tasks.
Experimental Thoroughness: 4/5. Seven scoring datasets, two model judges, and two restoration frameworks provide broad coverage; human truthfulness assessment and interaction-cost analysis remain limited.
Writing Quality: 3/5. The motivation and examples are clear, but some implementation details depend on an unavailable appendix; damaged formula extraction also limits precise verification in this reading.
Value: 4/5. Useful for perceptual evaluation systems requiring inspectable evidence and restoration guidance, but not an unconditionally stronger score regressor.