acceptodds
Under review as a conference paper at ICLR 2027

An Evolutionary Agentic Approach for Open-ended Image Quality Perception

Abstract

Generative models are rapidly expanding image quality assessment (IQA) beyond traditional fidelity factors to emerging dimensions such as physical plausibility and text-rendering correctness. However, existing IQA models rely on fixed definitions and heavy supervision, making them difficult to extend to open-ended perceptual dimensions. We identify holistic bias as an important limitation: when scoring an unseen dimension, models reuse generic quality priors, leading to scoring errors and rank inversion. To address this, we propose PACE (Perceptual Agentic Collaborative Evolution), a training-free multi-agent framework that formulates open-ended IQA as explicit protocol construction. Given a target dimension, PACE uses collaborative agents to construct an evaluation protocol composed of verifiable Visual Question Answering (VQA) probes, grounding evaluation in concrete visual evidence rather than holistic impressions. The resulting protocol is calibrated using only four human-annotated images per dimension, while a dual-track scoring mechanism aligns model perception with human scoring scales. Across traditional IQA, structural fidelity, context-aware aesthetics, and newly defined open-ended dimensions, PACE consistently improves its MLLM backbone, achieving competitive performance across diverse IQA settings, and reduces the Holistic Override Rate (HOR) from 44.4% to 8.6%.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Image Quality Assessment, Multimodal Large Language Model, Multi-Agent System

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_7gB7OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents PACE, a training-free framework for open-ended image quality assessment. Its central premise is that direct multimodal scoring can conflate overall visual quality with performance on a specific evaluation dimension. PACE addresses this problem by constructing explicit evaluation protocols through concept decomposition, visual question generation, and iterative critique. It combines protocol-based scoring with relative comparison against human-rated reference images. The framework is evaluated on conventional IQA, structural integrity, context-dependent aesthetics, and six newly constructed perceptual dimensions. The results show substantial improvements on several open-ended dimensions.

Critical strengths
  1. Adapting image evaluation to new, user-specified dimensions is an important problem. Formulating adaptation as explicit protocol construction provides a plausible way to make evaluation criteria more specific and inspectable.

  2. Promising empirical improvements over the same backbone. On the six open-ended dimensions, PACE improves average PLCC/SRCC from 0.493/0.501 to 0.714/0.666 over Qwen2.5-VL. The gains on text rendering, tactile satisfaction, and human vitality are encouraging.

  3. The paper includes multiple evaluation settings, component ablations, an additional backbone, and runtime measurements. The six-annotator study and reported inter-annotator agreement provide useful support for the human ratings.

Critical weaknesses
  1. Scope of adaptation and dependence on human calibration. The formulation conditions evaluation on a user-specified dimension and reuses a protocol initially grounded in one image. The claimed low human overhead depends on substantial reuse of calibrated dimensions. If evaluation requests frequently introduce genuinely new dimensions, human calibration would be required for nearly every request, even though annotators rate reference images rather than the test images themselves.

  2. The evidence for broad open-ended generalization remains limited. The central benchmark contains only 30 images per dimension, covers six author-selected dimensions, and uses a single image generator followed by manual filtering. It only provides limited evidence for robustness across arbitrary dimensions, image sources, and content distributions.

  3. The ablations do not establish the mechanism behind the reported reduction in holistic bias. Table 5 compares Track 1, Track 2, and their combination on conventional IQA benchmarks using PLCC/SRCC. However, the central holistic-bias claim concerns the open-ended Tier-4 setting and the HOR metric.

Questions for authors

Please see the weaknesses. I would consider raising my rating if the authors address the concerns.

7 comments

Thank you for recognizing the promising gains on open-ended dimensions. We hope our response will resolve your concerns.

W1 — Human calibration cost. In practical IQA, a quality dimension is typically applied to multiple images, rather than introducing an entirely new dimension for every evaluation request. PACE therefore requires only four candidate images to be rated once per new dimension; the resulting two anchors and evaluation protocol are then reused (Sec. 4.4). Even when new dimensions arise, this is substantially less costly than collecting a large MOS dataset and fine-tuning a dedicated scorer for each task.

Moreover, Track 1 alone provides much of the Tier-4 improvement without relative anchor comparison (PLCC/SRCC: 0.634/0.597 vs. direct scoring 0.493/0.501; see W3), while the anchors further align scores with human preferences.

W2 (Generalization beyond Tier-4). To evaluate generalization beyond our original six dimensions, we further tested PACE on 12 additional dimensions from EvalMuse-40K, which includes diverse prompts and images generated by multiple T2I models. The results compare direct scoring with PACE:

EvalMuse dimensionDirect PLCC / SRCCPACE PLCC / SRCC
Object Counting0.231 / 0.1980.344 / 0.368
Spatial Layout0.369 / 0.3930.384 / 0.412
Camera Viewpoint & Framing-0.160 / -0.098-0.051 / -0.027
Material & Medium0.254 / 0.1030.330 / 0.293
Action & Activity-0.049 / -0.0880.023 / -0.054
Animal & Character Integrity0.527 / 0.4040.555 / 0.467
Object Presence & Identity0.055 / 0.0020.358 / 0.323
Human Anatomy & Identity0.521 / 0.5660.535 / 0.639
Scene Context & Location0.420 / 0.3280.474 / 0.404
Color Attribute Fidelity0.002 / -0.0010.071 / 0.042
Named Artistic Style0.282 / 0.2490.396 / 0.288
Time & Weather Attributes0.530 / 0.4450.639 / 0.581
Macro average (12 dimensions)0.249 / 0.2080.338 / 0.311

Across all 12 dimensions, macro PLCC improves from 0.249 to 0.338, and macro SRCC from 0.208 to 0.311. Both correlations improve on all 12 dimensions, spanning diverse visual attributes and fine-grained generation requirements. These results further support PACE's generalization beyond the original Tier-4 evaluation set.

W3 (Tier-4 ablations and holistic bias). We added component ablations directly on the six open-ended Tier-4 dimensions, using the same evaluation setting:

ConfigurationPLCC ↑SRCC ↑
Direct Qwen2.5-VL0.4930.501
Track 1 Only0.6340.597
Track 2 Only0.5240.504
Without Critic0.6500.618
Full PACE0.7140.666

The results show that protocol-based evaluation is a major contributor: Track 1 alone substantially improves over direct scoring, while Critic refinement and dual-track fusion further improve the final correlations. Complementing these ablations, the original Q-Align vs. PACE evaluation shows that HOR decreases from 44.44% to 8.64% and BalancedHOR from 23.36% to 7.73% (see brPK W1), supporting the framework's effectiveness in reducing over-scoring on open-ended dimensions.

Thank you for the additional experiments. The evaluation and ablations address part of my concerns. Two questions remain:

  1. Although a quality dimension can be shared across images, what ensures that a protocol grounded in one image and its calibration anchors remain valid across diverse contexts? When is recalibration required?
  2. The ablations report PLCC/SRCC rather than HOR. Could you report component-wise HOR with explicit calibration settings to distinguish improvements from protocol-based evaluation versus reference comparison or score rescaling? These clarifications would strengthen the interpretation of the results.

Thank you for your timely reply and for raising these helpful questions. We hope the following results address your concerns.

Q1 — Protocol and anchor reuse

As described in Sec. 4.3, the initial image provides visual grounding, while the generated protocol captures dimension-level criteria rather than image-specific details, allowing reuse across images under the same dimension. For efficiency, PACE reuses existing protocols and anchors by default.

Anchors can be recalibrated when the quality range changes, while protocols are reconstructed when new visual contexts introduce criteria not covered by the original probes. This balances efficiency and adaptability without retraining.

Q2 — Component-wise HOR and calibration settings

We further computed HOR for the Tier-4 component ablations (GT ≤ 3), including an additional score-rescaling control:

ConfigurationHOR ↓
Direct Qwen2.5-VL38.3%
Direct + Anchor Rescaling34.6%
Track 1 Only11.1%
Track 2 Only16.0%
Without Critic9.9%
Full PACE8.6%

Calibration settings: Direct + Anchor Rescaling maps direct predictions to the human-rated anchor range without protocol evaluation or visual reference comparison. Track 1 Only scores directly on the original 1–5 scale, without anchor comparison or rescaling. Track 2 Only uses anchor-based comparison and score calibration, while Full PACE combines both tracks (Eq. 10).

Score rescaling alone reduces HOR only from 38.3% to 34.6%, whereas Track 1 reduces it to 11.1% without calibration. This confirms that protocol evaluation, rather than score rescaling alone, drives most of the HOR reduction. Reference comparison provides additional benefits, with Full PACE achieving the lowest HOR (8.6%).

Reviewer_brPKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper identifies "holistic bias" in MLLM-based IQA — models reuse generic quality priors on unseen dimensions — and proposes PACE, a training-free multi-agent framework that constructs per-dimension VQA evaluation protocols on the fly, with a one-time HITL calibration (four generated images as anchors) and dual-track scoring.

Critical strengths
  • The paper's core observation—that HOR and the quality dimensions in IQA evolve faster than benchmarks and annotations can keep pace with—is a valuable and well-founded insight.The evidence (e.g., Table 8: holistic scorers give ~3–4 to images with human GT 2.09) convincingly documents holistic bias.
  • The concept of on-the-fly, task-specific protocol construction is appealing; reusable and interpretable protocols are a real advantage over black-box scoring.
  • HOR is a dedicated, targeted metric for the failure cases on poor-quality samples, complementing PLCC/SRCC.
Critical weaknesses
  • HOR is one-directional and gameable. It only counts low-GT images receiving high scores. A model that scores everything low gets HOR = 0 yet is clearly not hallucination-free — so a low HOR does not certify correctness. Moreover, the reverse failure (high-quality images receiving low scores, i.e., false alarms) is never examined, even though PACE's projection onto [Y_low, Y_high] structurally invites under-scoring. Please report a symmetric error analysis.

  • The Planning Agent's decomposition is asserted, not proven. Why wouldn't the sub-dimensions themselves hallucinate? What guarantees completeness/accuracy on concepts the backbone barely knows? The problem definition may be fine, but who can play the planner role matters — there is no evidence that the backbone is qualified. A decomposition-quality study (expert review of completeness/accuracy) is missing. Also, is n ∈ [3,5] sufficient for real dimensions, and is there any ablation over n?

  • The "four-image" HITL calibration is under-defined and unvalidated. Why four? The anchors are synthesized by SD1.5, a weak generator — how can they be trusted, and validation should be reported. Notably, in all experiments the anchors are actually taken from dataset annotations (Sec. 5.1), so the claimed generative calibration pipeline is never quantitatively validated. Also, if a test image exceeds the high anchor in quality (e.g., from a different generator, seedance2.5), are the anchors still valid? Please ablate anchor count and validate the generative pipeline.

  • VQA probes may only detect a quality lower bound. Beyond some threshold, the MLLM can answer verifiable probes correctly regardless of quality, so probes stop discriminating good from excellent. This saturation in the high-quality regime is not discussed but directly limits the method's usefulness.

  • Visual evidence ≠ correctness; probe design stays traditional. Crop & Zoom targets traditional features (noise, sharpness) the agent already knows. For genuinely AIGC-specific issues (e.g., the "AI look"), how would such dimensions be defined and evaluated? The paper always reuses the same few examples (tactile satisfaction, human vitality); a broader set of concrete AIGC quality problem types would help readers judge real coverage.

  • The Slow pipeline is far weaker than specialized models on traditional tasks. On Tier 1, Slow is far below Fast (e.g., KADID 0.782 vs. 0.926). Beating the Qwen2.5-VL baseline is expected. The key promise of this methodology is that, on genuinely unseen problems, the Slow pipeline approaches dedicated SOTA built with full annotations — this is not yet shown.

  • Routing granularity is underspecified. How fine-grained is dimension similarity, and how was θ = 0.8 chosen? Semantically close but perceptually distinct dimensions (e.g., "tactile satisfaction" vs. "surface smoothness") could be conflated. A mis-routing analysis is needed.

Questions for authors

please answer the questions in "Weaknesses" part.

3 comments

Thank you for highlighting the value of reusable, interpretable evaluation protocols. We hope our response will resolve your concerns.

W1 — Asymmetric errors. HOR measures over-scoring of low-quality images. We additionally evaluated ReverseHOR = P(pred ≤ 3 | GT ≥ 4) and BalancedHOR = (HOR + ReverseHOR)/2 on the original Q-Align/PACE predictions (Fig. 4):

Metric ↓Q-AlignPACE
HOR44.44%8.64%
ReverseHOR2.27%6.82%
BalancedHOR23.36%7.73%
MAE1.2000.701
Mean signed bias+1.029+0.251

The substantial reductions in BalancedHOR and MAE show that PACE's lower HOR is not achieved by simply shifting predictions downward. For GT-low images, Q-Align scores 3.75 on average versus human MOS 2.09 (Table 8).

W2 — Planner reliability. PACE does not rely on a single unconstrained decomposition. The Planning Agent identifies visually observable sub-dimensions, the Visualizer constructs corresponding VQA probes, and the Critic iteratively refines and corrects these probes (Sec. 4.3; Fig. 6). An additional Human-based decomposition audit yielded 91% correctness and 96% completeness, supporting the reliability of the generated sub-dimensions. We also varied the number of sub-dimensions:

Number of sub-dimensions357
PLCC.616.743.675
SRCC.595.693.625

Five performs best, supporting the 3–5 range as a balance between concept coverage and redundant criteria.

W3 — HITL calibration. The key is a pair of human-rated quality anchors, rather than SD1.5 itself. Generating four candidates provides lightweight redundancy for selecting high- and low-quality references, while the final scoring requires only two anchors. Candidates can come from SD1.5, stronger generators, or annotated datasets. Humans rate them, and the highest- and lowest-rated are retained (Sec. 4.4). Our benchmark uses dataset anchors to evaluate the scoring mechanism (Sec. 5.1). We tested the number of scoring anchors:

Scoring anchors125
Mean PLCC ↑.615.652.647
Mean SRCC ↑.583.612.610

Two anchors perform best, supporting the lightweight calibration design. If a test image exceeds the upper anchor in quality, the anchor can be replaced with a higher-quality human-rated reference without retraining. We further discuss high-quality discrimination in W4. We also clarify that Seedance 2.5 is a video generator, whereas our task concerns still-image IQA.

W4 — High-quality saturation. High-quality saturation is a general challenge in IQA: even supervised scorers may struggle to distinguish quality differences beyond the range represented by their MOS training data. PACE offers greater flexibility through dual-track scoring: Track 2 provides relative comparisons against human-rated anchors, complementing potentially saturated VQA probes in Track 1. PACE can replace the upper anchor with a higher-quality human-rated reference without retraining (Eq. 10). This allows the relative comparison and scoring range to adapt as image quality improves, providing complementary discrimination when VQA probes become less informative.

W5 — AIGC-specific dimensions. PACE is not restricted to traditional IQA features. For an AIGC-specific concept such as AI look, observable criteria could include unnatural skin smoothing, repetitive textures, and inconsistent material reflections. The Planning Agent defines the target-specific criteria, while Crop & Zoom only assists visual inspection. Our six Tier-4 dimensions cover several such generation-specific concerns (Ue5w W4), and the additional EvalMuse experiments extend the evaluation to further visual attributes (7gB7 W2).

W6 — Slow versus supervised IQA. Fast and Slow address different evaluation regimes. We agree that Slow is weaker than specialized supervised IQA models on established benchmarks. Its purpose is not to replace those specialists, but to construct evaluators for newly requested dimensions without dedicated MOS training data. When labeled data become available, trained specialists can be incorporated into Fast, making the two paths complementary.

W7 — Routing threshold. Routing accelerates protocol reuse, while unseen dimensions trigger Slow construction. We evaluated 48 queries covering canonical names, paraphrases, descriptions, and perceptually distinct near-confusers. θ = 0.8 achieves the highest routing accuracy (77.1%), compared with 66.7% at θ = 0.7 and 68.8% at θ = 0.9. At θ = 0.8, only 3 queries are incorrectly routed to Fast, compared with 1 at θ = 0.9, while unnecessary Slow calls decrease from 14 to 8. Thus, θ = 0.8 provides a practical balance between incorrect protocol reuse and routing efficiency.

Dear Reviewer,

Thank you again for your thoughtful and constructive feedback on our submission. We sincerely appreciate your time and effort throughout the review process.

As the November 18 discussion deadline approaches, we would like to kindly follow up on our previous responses and additional experiments. We hope they have adequately addressed your concerns.

If you have had a chance to review our responses, we would greatly appreciate knowing whether any concerns remain unresolved. We would be more than happy to provide further clarification or additional information before the discussion period ends.

Thank you once again for your time and consideration.

Best regards,

The Authors of Paper 3073

Dear Reviewer,

Thank you again for your thoughtful and constructive feedback on our submission. We sincerely appreciate your time and effort throughout the review process.

As the November 18 discussion deadline approaches, we would like to kindly follow up on our previous responses and additional experiments. We hope they have adequately addressed your concerns.

If you have had a chance to review our responses, we would greatly appreciate knowing whether any concerns remain unresolved. We would be more than happy to provide further clarification or additional information before the discussion period ends.

Thank you once again for your time and consideration.

Best regards,

The Authors of Paper 3073

Reviewer_Ue5wOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes PACE, a multi-agent framework for open-endeed image quality assessment. The paper's main idea is to break down a target quality dimension into several VQA questions, refine them using multiple agents, and then combine the VQA results with reference-image comparisons to produce a final score. The paper also introduced HOR to measure the effect of holistic bias.

Critical strengths

The motivation seems practical. MLLMs basd IQA may be affected by overall image quality when evaluating a specific dimension.

Critical weaknesses

W1. Limited novelty. The paper lacks clear algorithmic contributions. Most of the proposed method is an engineering combination of existing techniques rather than a new algorithm. The overall framework also lacks theoretical justification for its design choices.

W2. The main experiments only use Qwen2.5-VL-7B, which is not a sota mllm. Given that PACE is a training-free framework, I don't see why it cannot be evaluated on more recent and stronger MLLMs. One of the main claims is that existing MLLMs suffer from holistic bias, but it is unclear whether this issue is still as severe with stronger models.

W3. Most of the ablation experiments are conducted on traditional IQA datasets, rather than the open-ended dimensions that motivate the paper.

W4. The choice of the six open-ended dimensions seems rather arbitrary. The authors claim that these dimensions are representative, but they don't really explain why these six were selected or what makes them representative.

Questions for authors

see weaknesses

4 comments

Thank you for recognizing the practical motivation for dimension-specific IQA. We hope our response will resolve your concerns.

W1 (Novelty and design rationale). PACE constructs and calibrates an IQA evaluator for an unseen dimension at inference time, rather than training a dimension-specific scorer. Our methodological contribution lies in dynamically constructing reusable evaluation protocols for previously unseen perceptual dimensions. Existing IQA approaches typically rely on MOS training data for predefined quality dimensions. PACE instead decomposes a user-specified dimension into observable factors and iteratively refines them into VQA probes (Sec. 4.3). Track 1 aggregates probability-weighted probe scores (Eqs. 4–6), while Track 2 compares the image with human-rated quality anchors (Eqs. 7–8); their fusion produces a continuous, human-aligned score (Eqs. 9–10).

This design is motivated by rubric/checklist-based LLM evaluation, where explicit criteria improve the reliability and interpretability of subjective judgments (e.g., LLM-Rubric and CheckEval). In IQA, making the requested criteria explicit reduces interference from the MLLM's generic quality priors, while lightweight anchor calibration avoids collecting a new MOS training set for each dimension.

W2 (Stronger MLLMs and holistic bias). We evaluated five additional backbones on Tier-4, including larger Qwen models and InternVL3.5. Direct scoring still assigns high scores to many low-GT images: HOR ranges from 19.8% to 42.0%. PACE reduces HOR on every model, while improving both PLCC and SRCC.

BackbonePLCC (Direct → PACE) ↑SRCC (Direct → PACE) ↑HOR (Direct → PACE) ↓
Qwen3-VL-8B.602 → .727.607 → .67542.0% → 11.1%
Qwen3.5-9B.634 → .723.657 → .67539.5% → 1.2%
Qwen3-VL-32B.649 → .712.617 → .65733.3% → 8.6%
Qwen3.5-27B.662 → .684.623 → .64419.8% → 3.7%
InternVL3.5-14B.561 → .642.613 → .63340.7% → 21.0%

These results show that the over-scoring failure is not specific to Qwen2.5-VL-7B. Stronger backbones do not eliminate this issue, and dimension-specific protocols remain beneficial across all five models.

W3 (Ablations on open-ended dimensions). We added component ablations on all six Tier-4 dimensions under the same evaluation setting:

ConfigurationPLCC ↑SRCC ↑
Direct Qwen2.5-VL0.4930.501
Track 1 Only0.6340.597
Track 2 Only0.5240.504
Without Critic0.6500.618
Full PACE0.7140.666

Track 1 provides the main gain in dimension-specific ranking, while Track 2 contributes human-scale calibration. The improvement from Without Critic to Full PACE further supports iterative protocol refinement, and the full dual-track framework achieves the highest overall correlations.

W4 (Choice of six dimensions). We selected six dimensions to cover complementary AIGC quality concerns: textual correctness (Typo), physical consistency (Light), material perception (Tactile), human naturalness (Portrait), and creative coherence and storytelling (Dream and Cine). These represent distinct evaluation challenges that can be overlooked by holistic image scores and are also reflected in the evaluation criteria of Qwen-Image-Bench (QIB):

Our dimensionRelated QIB evaluation criteria
TypoText Accuracy and Text Layout: correctness and organization of rendered text
LightPhysical Logic and Lighting & Atmosphere: plausible shadows, reflections, and illumination
TactileMaterial Texture, Material Properties, and Naturalness: visual cues underlying perceived touch
DreamImagination and Feature Matching: coherent integration of surreal or mixed elements
PortraitFacial Expression, Anatomical Fidelity, and Emotional Expression: natural human appearance and expression
CineCinematic Style, Camera / Lens Style, and Composition: framing, style, and visual composition

Thanks for the additional experiments, which address some of my concerns, but I'm still not fully convinced by paper's motivation and novelty. The results show that PACE improves performance, but it's still unclear whether holistic bias is really the main cause of the errors. Also, the overall method still feels like a combination of existing ideas, with limited algorithmic novelty. Overall, I will keep my rating unchanged.

Thank you for your timely reply. We provide a comparison with representative IQA methods and further examine the relationship between holistic aesthetics and dimension-specific scoring errors. We hope these additional analyses address your remaining concerns and encourage you to reconsider your rating.

Q1 — Methodological comparison

Probability-weighted scoring has been explored in prior IQA methods (e.g., Q-Align). To our knowledge, PACE is the first framework to jointly (1) dynamically construct, validate, and reuse an IQA evaluator for previously unseen perceptual dimensions through multi-agent interaction at inference time, and (2) integrate lightweight, one-time human-in-the-loop (HITL) calibration into this process to align scores with human preferences, without dimension-specific model training.

MethodTraining-freeUnseen-dimension ScoringMulti-agent Protocol ConstructionIterative Criterion RefinementOne-time HITL CalibrationProtocol Reuse
Q-Align××××××
DeQA-Score××××××
Q-Scorer××××××
Compare2Score××××××
Co-Instruct××××××
Dog-IQA✓×××××
Direct MLLM✓✓××××
PACE (Ours)✓✓✓✓✓✓

✓ indicates that the method explicitly supports the corresponding capability; × indicates that the capability is not an explicit component of the original method. “Unseen-dimension scoring” refers to scoring previously unseen, user-specified perceptual quality dimensions. “One-time HITL calibration” refers to rating candidate reference images once for each new dimension and retaining two human-rated anchors for subsequent scoring.

Q2 — Overall aesthetics versus dimension-specific scores

We additionally measured the correlation between overall aesthetic quality (reference labels combining human and Gemini assessments) and dimension-specific scores:

Dimensiondimension GTQ-AlignPACE (seed average)
Typo+0.439+0.612+0.497
Light−0.002+0.221+0.214
Tactile+0.384+0.586+0.397
Dream+0.386+0.556+0.472
Vitality+0.163−0.127+0.071
Cine+0.346+0.696+0.447
Macro mean (six dimensions)+0.286+0.424+0.350

Each value is the correlation with the same overall-aesthetics reference. Macro means are unweighted averages across the six displayed dimensions.

Q-Align is more strongly correlated with overall aesthetics than human dimension GT in five of six dimensions. More importantly, PACE's aesthetic-correlation pattern is closer to human dimension GT in all six, reducing the mean absolute correlation gap from 0.235 to 0.094. Together with the HOR results and Table 8, these findings provide further evidence that holistic aesthetic priors interfere with dimension-specific quality assessment, and that PACE reduces this interference.

Dear Reviewer,

Thank you again for your thoughtful and constructive feedback on our submission. We sincerely appreciate your time and effort throughout the review process.

As the November 18 discussion deadline approaches, we would like to kindly follow up on our previous responses and additional experiments. We hope they have adequately addressed your concerns.

If you have had a chance to review our responses, we would greatly appreciate knowing whether any concerns remain unresolved. We would be more than happy to provide further clarification or additional information before the discussion period ends.

Thank you once again for your time and consideration.

Best regards,

The Authors of Paper 3073

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.