acceptodds
Under review as a conference paper at ICLR 2027

RLA-Net: Rubric-Grounded Learning for Decomposable Image Aesthetic Assessment

Abstract

A single aesthetic score summarizes preference but does not explain the visual strengths, weaknesses, and trade-offs behind it. Learning interpretable image aesthetic assessment (IAA) from aggregate ratings is challenging: the ratings leave individual aesthetic states underdetermined, prompt–feature similarity does not ensure attribute-specific visual responses, and rubric states alone do not specify their image-dependent importance. We introduce RLA-Net, a CLIP-based framework that jointly estimates readable aesthetic states and learns their adaptive scoring roles through rubric-grounded score composition. Local rubric grounding estimates grades by matching visual evidence to three language-defined reference states for each of eight aesthetic dimensions. Contrastive Rubric Alignment (CRA) complements overall-score supervision with source-relative, rubric-targeted synthetic pairs that constrain global image–anchor responses in prescribed directions. Gated Aesthetic Scoring combines rubric grades, visual descriptors, and image context into explicit prediction terms, allowing similar states to play different scoring roles across images. Training uses image-level ratings, fixed rubric anchors, and synthetic pairs, without per-image rubric annotations, comments, critiques, or teacher-generated targets. Experiments on AVA, PARA, TAD66K, and FLICKR-AES demonstrate competitive performance.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Image Aesthetic Assessment, Vision-Language Models, Interpretable Visual Aesthetic Evaluation

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

This paper proposes RLA-Net, a CLIP-based framework for decomposable image aesthetic assessment. It models aesthetic scores through eight language-defined rubrics and introduces Contrastive Rubric Alignment (CRA) to learn rubric-specific directions from synthetic perturbations, together with Gated Aesthetic Scoring for adaptive rubric aggregation. Experiments on AVA, PARA, TAD66K, and FLICKR-AES show competitive prediction performance, and ablations demonstrate the effectiveness of CRA and routing.

Critical strengths

(1) The paper goes beyond single-score prediction by explicitly modeling rubric states and their contributions, providing a more inspectable aesthetic assessment framework. (2) CRA provides additional rubric-level directional signals without requiring per-image attribute annotations, and improves prediction performance.

Critical weaknesses

(1) Although interpretability is a central motivation, the paper does not verify whether learned rubric states and routing weights agree with human aesthetic judgments. The results show predictive utility but not necessarily human-aligned explanations. (2) Some perturbations affect multiple rubrics simultaneously (e.g., blur affects sharpness and texture). The paper does not sufficiently demonstrate that individual rubric scores capture distinct aesthetic factors. (3) The paper does not compare with MLLM-based aesthetic assessment methods, although they are relevant to the interpretability motivation. Please clarify why these methods are excluded and how RLA-Net differs from them beyond computational efficiency.

Questions for authors

Q1. Have the authors evaluated whether predicted rubric states or routing weights correlate with human judgments of aesthetic factors? Q2. How does the model prevent correlated perturbations from causing entangled rubric representations (e.g., sharpness vs. texture)? Q3. Have the authors considered more fine-grained rubric anchors beyond three levels?

4 comments
AuthorsOpenReview ↗

Response to Reviewer yjRA - Part I (W1,W2,Q1,Q2)

Response Summary: We are grateful for your careful reading and constructive feedback. We appreciate your recognition of our inspectable rubric-based scoring and the predictive benefits of CRA’s directional guidance without per-image attribute annotations. Your questions have helped us strengthen the evaluation of human alignment and rubric specificity, and clarify our positioning relative to MLLM-based methods. Below, we respond to each weakness and question point by point, supported by additional analyses and experiments. All new results and experimental details are included in the revised PDFs.

W1 and Q1: Alignment with human aesthetic judgments.

We agree that overall-score accuracy alone does not establish human-aligned rubric states. We therefore added two complementary evaluations: public benchmarks for comparison against independently collected human annotations, and an expert-annotated dataset covering all eight rubrics. All methods follow the same zero-shot attribute-evaluation protocol, without additional training or fine-tuning on these evaluation annotations.

1. Public human-annotated benchmarks. We evaluate 393 EVA images and 3,000 PARA images. Their annotation schemes do not match our full eight-rubric system, so we compare to the available, semantically overlapping criteria related to lighting, color, and composition. All methods are evaluated on the same images and reference annotations.

MethodEVA MAEEVA RMSEPARA MAEPARA RMSE
AesCLIP-A0.2500.2890.3380.369
AesCLIP-B0.2240.2650.2900.325
AesCLIP-Ens0.2070.2480.2810.314
RLA-Net0.1500.1720.1080.134

2. Expert annotations covering all eight rubrics. To complement the partial attribute coverage of public benchmarks, we collected annotations for 100 images from 20 aesthetic experts, using criteria semantically aligned with our eight rubrics. We compare all methods under the same zero-shot evaluation protocol, using these annotations exclusively for evaluation.

MethodMAEMacro RMSE
AesCLIP-A0.2440.296
AesCLIP-B0.2430.299
AesCLIP-Ens0.2200.271
RLA-Net0.2050.241

RLA-Net achieves the lowest MAE and macro RMSE in all three evaluations. The public benchmarks provide evidence of human alignment for overlapping attributes, while the expert study extends this assessment to the complete eight-rubric system. These results directly evaluate rubric-state agreement with human judgments, complementing overall-score accuracy and grade-distribution analyses. More detailed information is provided in Section E.5, Table E.5, and Table E.6 of the updated supplementary PDF.

W2 and Q2: Rubric specificity under correlated perturbations.

Thank you for this question. RLA-Net defines each rubric through distinct language anchors, such as "sharp focus" and "rich texture," with each grade computed over its own rubric levels. CRA constrains the targeted rubric's global image-anchor response; it neither assigns a common degradation target to all rubrics nor requires the others to remain unchanged. Shared encoders and anchors connect this guidance to local grounding during joint training. These mechanisms promote rubric-specific responses while allowing related attributes to change together; they do not impose statistical independence.

We test this on two AVA samples (492626 and 795095), comparing each original with global blur, affecting edge clarity and surface detail; edge softening, primarily affecting sharpness; and surface smoothing, primarily affecting texture. The table reports rubric scores and changes from the originals.

ImageEditSharpnessΔ SharpnessTextureΔ Texture
492626Original0.839-0.678-
Global blur0.774-0.0650.566-0.112
Edge softening0.809-0.0300.681+0.004
Surface smoothing0.842+0.0030.623-0.055
795095Original0.871-0.726-
Global blur0.846-0.0250.676-0.050
Edge softening0.859-0.0120.730+0.003
Surface smoothing0.873+0.0020.702-0.025

Results. Global blur lowers both scores in both examples. Edge softening lowers sharpness by 0.030/0.012, while texture changes by at most 0.004; surface smoothing lowers texture by 0.055/0.025, while sharpness changes by at most 0.003. Thus, the rubrics exhibit both coordinated responses to shared degradation and differentiated responses to attribute-focused edits. These examples support distinguishable rubric responses consistent with the intended visual changes, rather than an indiscriminate shared degradation score. Visual comparisons are provided in Section E.9, Fig. E.2 of the updated supplementary PDF.

AuthorsOpenReview ↗

Response to Reviewer yjRA - Part II (W3, Q3)

W3. Comparison with MLLM-based aesthetic assessment.

Thank you for suggesting this comparison. As RLA-Net is a CLIP-based model, our original evaluation prioritized closely related methods. MLLM-based methods differ substantially in model scale and computational requirements, and their supervision also varies: ArtiMuse includes aesthetic-caption pretraining [2], whereas Q-Align derives textual rating levels from existing scores [1]. We appreciate the value of a broader comparison and have therefore added both methods:

MethodTotal params (B)AVA SRCCAVA PLCCPARA SRCCPARA PLCC
Q-Align [1]≈8.20.8220.8170.9230.940
ArtiMuse [2]≈8.00.8270.8260.9360.958
RLA-Net (ours)0.4340.8130.8120.9270.951

With 0.434B parameters, approximately 5% of either MLLM's parameter count, RLA-Net achieves competitive correlations: it exceeds Q-Align on PARA and remains within 0.014 of ArtiMuse on both metrics across the two benchmarks. These results are obtained without additional aesthetic-caption or critique supervision. Beyond model size, RLA-Net provides an explicit rubric-conditioned scoring process, jointly learning readable attribute states and their image-dependent use in prediction.

References

[1] Wu et al. Q-Align: Teaching LMMs for Visual Scoring via Discrete Text-Defined Levels. ICML, 2024.

[2] Cao et al. ArtiMuse: Fine-Grained Image Aesthetics Assessment with Joint Scoring and Expert-Level Understanding. CVPR, 2026.

Q3. Finer-grained rubric anchors beyond three levels.

Thank you for this suggestion. We have extended the anchor-granularity ablation to two, three, and five levels on AVA and PARA. These anchors define semantic reference states rather than discrete output categories: soft assignments to the anchors yield continuous rubric scores.

No. LevelsAVA SRCCAVA PLCCPARA SRCCPARA PLCC
20.7940.7930.9020.924
3 (default)0.8130.8120.9270.951
50.7100.7150.8220.865

Three levels achieve the highest correlations on both datasets. Increasing from three to five levels reduces SRCC/PLCC by 0.103/0.097 on AVA and 0.105/0.086 on PARA, while two levels also perform below the default. Thus, finer-grained anchors do not improve prediction in the tested setting. These results support retaining three levels as a compact semantic reference system that still permits continuous rubric estimates. The additional results and anchor definitions are provided in Section E.11, Table E.12 of the updated supplementary PDF.

Closing: Thank you again for your time, effort, and constructive feedback; we welcome any follow-up questions and will respond promptly.

I appreciate the authors' detailed rebuttal and the substantial additional experiments. The new human-annotated evaluations, MLLM comparisons, and extended ablations address several of my original concerns.

However, I share Reviewer zD3u's concern that the interpretability claim remains insufficiently validated. While the rubric grades show improved alignment with human judgments, the faithfulness of their scoring contributions remains unclear. Moreover, the perturbation analysis on only two images provides limited evidence of rubric specificity.

I am inclined to maintain my initial rating.

AuthorsOpenReview ↗

Response to Reviewer yjRA

Thank you for the follow-up and for recognizing the human-annotated evaluations, MLLM comparisons, and extended ablations. The human validation in Section E.5 establishes that the rubric grades agree with human judgments.

Regarding the faithfulness of the rubric scoring contributions, we would kindly note that this question was also discussed in our response to Reviewer zD3u, including exact score reconstruction, grade-only readouts, and deletion and edit tests (Sections E.6–E.7). We truly appreciate you raising this point again, as it encouraged us to examine it more closely. We have therefore conducted additional experiments that further strengthen this evidence, and extended the specificity analysis beyond two images. New results are in Sections E.9, E.12, and E.13 of the updated supplementary PDF.

(1) Faithfulness of scoring contributions.

Because the prediction is an exact sum of the rubric terms, we measure the role of the grades directly by intervening on the inputs of the frozen scoring head.

Pathway exchange (Section E.13, Table E.14). For each original–edited image pair, we cross the eight grades and the high-dimensional features of the two images; the signed effects of the two pathways sum to the full logit change.

EditFull changeGrade pathwayFeature pathway
Global blur0.5500.2400.330
Desaturation0.0770.0420.031
Luminance flattening0.2800.1450.145
Brightening0.2610.1300.142
Surface smoothing0.0330.0180.017
Edge softening0.0670.0570.059

For all six edits, the grade pathway carries 44–85% of the full change, comparable to or larger than the high-dimensional features.

Single-grade replacement (Table E.15). Replacing one named grade at a time, we compare the logit change caused by the targeted rubric with the mean change caused by the other seven.

Edit / target rubricTargetOther sevenPaired difference
Desaturation / color0.03120.0043+0.0269
Luminance flattening / contrast0.09570.0129+0.0828
Brightening / exposure0.08660.0131+0.0735
Surface smoothing / texture0.01320.0025+0.0107
Edge softening / sharpness0.00490.0045+0.0004

For all five edits, the targeted grade changes the score more than the other seven grades on average. For color, contrast, exposure, and texture, the target effect is 5–7 times the seven-rubric mean. Each named grade thus acts on the score in correspondence with its own attribute.

Removing the global residual (Table E.16). With the weights of the full model fixed, we remove only the global residual FzF_z from the shared context and evaluate on all 19,929 AVA test images:

SettingSRCCPLCCMAE
Full model0.8130.8120.0381
Global residual removed0.7520.7620.0562

The model retains 92% of its ranking performance, so the prediction does not rest primarily on this bypass. Removing the global residual lowers SRCC to 0.752, whereas removing the grade residual lowers it to 0.648 (Table E.7), so the prediction depends more on the named grades than on this bypass.

(2) Rubric specificity beyond two images.

Ten additional images and three rubric pairs (Figures E.3–E.5). Across sharpness/texture, color/contrast, and exposure/contrast, every joint edit lowers both grades on all twelve images, while every targeted edit lowers its own grade by 0.012–0.080 and changes the paired grade by at most 0.005.

The full AVA test set (Section E.12, Table E.13). On all 19,929 test images:

Pair (A / B)Targeted editMean Δ targetedMean Δ paired
Sharpness / TextureEdge softening → sharpness−0.0236+0.0045
Surface smoothing → texture−0.0477+0.0026
Color / ContrastDesaturation → color−0.0767+0.0028
Luminance flattening → contrast−0.0363+0.0039
Exposure / ContrastBrightening → exposure−0.0740+0.0028
Luminance flattening → contrast−0.0428+0.0038

Each targeted edit changes its own grade 5–27 times more than the paired grade. The pattern holds image by image: for the middle half of the test set, every targeted grade decreases by at least 0.0214. Joint edits, which degrade both attributes, lower both grades (Table E.13).

RLA-Net's interpretability is now supported at three levels: the grades agree with human attribute annotations (Section E.5), they carry a substantial and rubric-specific share of the scoring computation (Section E.13), and they respond selectively to their own attributes across three rubric pairs and the full AVA test set (Sections E.9 and E.12). We hope these results resolve the remaining concerns, and we would be grateful if you would take them into account in your final assessment.

Reviewer_zD3uOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes RLA-Net for image aesthetic assessment from aggregate image-level ratings. Its motivating distinction is between a rubric state, such as an image's lighting or composition, and that rubric's image-dependent role in the final aesthetic score. Eight rubrics are represented by three language anchors each. Patch-anchor similarities produce local rubric grades and descriptors. Contrastive Rubric Alignment (CRA) uses synthetically perturbed source-image pairs to push the target rubric's global anchor response in a prescribed direction. A gated cross-attention scoring module then combines rubric grades, rubric descriptors, and global image context into eight additive pre-sigmoid prediction terms.

The method is trained without per-image rubric labels, critiques, or teacher targets, using overall ratings, hand-specified language anchors, and synthetic pairs. RLA-Net reports SRCC/PLCC of 0.813/0.812 on AVA, 0.927/0.951 on PARA, 0.508/0.539 on TAD66K, and 0.812/0.832 on FLICKR-AES. AVA ablations show that the complete model outperforms tested variants without CRA, with global rather than local grounding, with two anchors, or with InfoNCE. On seven non-mood rubrics over 3,000 AVA images, the supplement also reports exploratory output distances to the model-generated ArtiMuse-F reference: RLA-Net obtains MAE/macro RMSE of 0.1581/0.1830 versus 0.2757/0.3334 for AesCLIP-Ens.

The problem statement is clear and worthwhile, and the distinction between state and scoring role makes a coherent story. However, apart from one explicitly qualified exploratory comparison using a model-generated reference and unmatched output definitions, the experiments primarily validate overall MOS prediction rather than the central semantic claim that the exposed quantities are correct or faithful rubric states and roles. The paper itself carefully acknowledges that its computation does not establish human-aligned or causal explanations.

Critical strengths
  1. The paper identifies a meaningful limitation of scalar aesthetic scoring and articulates a clear distinction between rubric state and scoring role. The proposed components correspond logically to this decomposition, and the complete pre-sigmoid score is exactly reconstructable from output bias b0 and the rubric terms.

  2. Overall-score prediction is competitive. On AVA, RLA-Net's 0.813 SRCC exceeds the strongest listed baseline by 0.025, and the full model reaches 0.813/0.812 versus 0.794/0.793 without CRA. The paper is also commendably careful not to overstate every table: it notes the near-tie on PARA, the lower TAD66K ranking result, the mismatch between starred official-checkpoint tests and target-dataset training, and the limits of non-causal explanations. The supplement is similarly explicit about what the ArtiMuse-F comparison can and cannot establish.

Critical weaknesses
  1. The main claimed contribution, interpretable rubric-grounded decomposition, remains insufficiently validated. Aggregate-score accuracy and component ablations only show that the architecture helps prediction. Because descriptor features and context containing a global residual provide high-dimensional paths that do not pass through the scalar grades, accurate MOS predictions can coexist with arbitrary, collapsed, or non-human rubric states. Figure 3 displays routing weights rather than the complete additive terms, and its displayed grades occupy only a narrow range; Appendix B then compares semantically different post-processed outputs. The current evidence therefore establishes neither semantic state accuracy nor faithful scoring roles. Addressing this concern requires human rubric-level ground truth or rigorous intervention/faithfulness evidence.

  2. The empirical comparison and attribution remain insufficiently controlled for the quantitative claims. Table 2 mixes cited results, target-dataset training, and direct tests of official checkpoints without target fine-tuning; the large FLICKR-AES advantage over the strongest starred checkpoint is therefore not a matched method comparison. The AVA main result and ablations use a single seed, 42, without multi-seed uncertainty estimates. Routing and gating are changed jointly rather than isolated. The supplement now supplies detailed losses, preprocessing, pair construction, and ablation definitions, materially improving reproducibility, but it does not resolve these comparison and attribution issues or provide equally detailed training and model-selection settings for the other three datasets.

Questions for authors
  1. Beyond the model-generated ArtiMuse-F reference, can the authors quantitatively validate the eight grades and scoring roles against human rubric annotations or a convincing controlled benchmark? Please report full-dataset distributions and uncertainty for the original MOS-path grades, agreement with human states, robustness to semantically equivalent prompts, and counterfactual/intervention tests. Strong evidence here could materially change my rating.

  2. How much prediction is actually mediated by the named scalar grades rather than the global-context and high-dimensional descriptor bypasses? Please report actual additive terms t_c, grade interventions, completeness/sufficiency or deletion tests, and variants that remove F_z from the shared context and restrict the readout to rubric grades. Please also test whether text-anchor semantics are preserved after textual LoRA.

3 comments
AuthorsOpenReview ↗

Response to Reviewer zD3u - Part I (Q1)

Response to Reviewer zD3u

Thank you for the thorough review and thoughtful suggestions. We appreciate your recognition of the problem's importance, the distinction between rubric states and their scoring roles, and RLA-Net's competitive predictive performance. Your feedback has helped us clarify our claims and strengthen the experimental evaluation. Below, we respond to each weakness and question point by point. All new results are included in the updated supplementary PDF.

Q1. Human alignment and semantic stability of rubric grades.

Thank you for this question. In summary, we have added four complementary analyses of human alignment, full-test-set distributions, prompt robustness, and counterfactual responses to assess the semantic validity and stability of the rubric grades.

1. Human-annotated benchmarks. We assess human alignment through two complementary evaluations: public benchmarks with overlapping attribute annotations and an expert-annotated dataset covering all eight rubrics. Public annotations overlap with only part of our rubric system. We evaluate 393 EVA images using light-color and composition-depth labels, and 3,000 PARA images using lighting and composition labels. All RMSE values below are macro RMSE; lower values are better.

MethodEVA MAEEVA RMSEPARA MAEPARA RMSE
AesCLIP-A0.2500.2890.3380.369
AesCLIP-B0.2240.2650.2900.325
AesCLIP-Ens0.2070.2480.2810.314
RLA-Net0.1500.1720.1080.134

To assess all eight rubrics, we additionally evaluate 100 images annotated by 20 aesthetic experts using semantically aligned rubric definitions. These annotations are also reserved for evaluation, under the same protocol for all methods.

MethodMAEMacro RMSE
AesCLIP-A0.2440.296
AesCLIP-B0.2430.299
AesCLIP-Ens0.2200.271
RLA-Net0.2050.241

RLA-Net obtains the lowest errors in all three evaluations, providing human-reference evidence for rubric-state alignment beyond overall-score accuracy. Annotation and evaluation details are in Section E.5, Tables E.5 and E.6 in the supplementary PDF.

2. Full-test-set distributions. We provide per-rubric histograms and summary statistics over 19,929 AVA test images. The means and medians are summarized below, with full distributions and detailed statistics reported in Section E.2, Table E.2 and Fig. E.1 in the updated supplementary PDF.

StatisticLightingCompositionColorSharpnessContrastExposureTextureMood
Mean0.450.480.710.140.420.340.500.48
Median0.450.470.720.120.400.240.490.44

3. Semantically equivalent prompts. On 1,024 randomly selected images, we test rubric-grade consistency after anchor paraphrasing. Without templates, we directly encode phrases such as “bright lighting.” With templates, we encode each phrase and two photography-related variants, such as “a photo with bright lighting,” then normalize each embedding, average the three, and renormalize the result.

Anchor encodingMacro SRCC
Direct phrase encoding0.945
Phrase and template ensemble0.958

Both configurations retain strong rank agreement with the original grades under the tested prompt variations, with further improvement from template ensembling (Section E.7 of the updated supplementary PDF).

4. Counterfactual image edits. We test three edits on each of two AVA images: global blur affects both edge clarity and surface detail, while edge softening and surface smoothing primarily target sharpness and texture, respectively. The table reports rubric scores and changes from the originals to examine shared and selective responses.

ImageEditSharpnessΔ SharpnessTextureΔ Texture
AVA 492626Original0.839-0.678-
Global blur0.774-0.0650.566-0.112
Edge softening0.809-0.0300.681+0.004
Surface smoothing0.842+0.0030.623-0.055
AVA 795095Original0.871-0.726-
Global blur0.846-0.0250.676-0.050
Edge softening0.859-0.0120.730+0.003
Surface smoothing0.873+0.0020.702-0.025

Global blur lowers both rubric scores in both examples. Edge softening reduces sharpness by 0.030/0.012, with texture changes of at most 0.004; surface smoothing reduces texture by 0.055/0.025, with sharpness changes of at most 0.003. These examples illustrate coordinated responses to shared degradation and selective responses to attribute-focused edits. Visual comparisons are provided in Fig. E.9 of the updated supplementary PDF.

AuthorsOpenReview ↗

Response to Reviewer zD3u - Part II (Q2, W1)

Q2. Contribution completeness and predictive use of rubric grades. We added four tests in response to Q2.

1. Complete prediction terms. Figure 3 displays routing weights αc\alpha_c, whereas the complete terms tct_c in Eq. (8) also include contextual readouts and grade residuals. On 19,929 AVA test images, z=b0+∑ctcz=b_0+\sum_c t_c reconstructs the pre-sigmoid prediction with a maximum absolute error of 2.38×10−72.38\times10^{-7} (Section E.6). This verifies the numerical completeness of the pre-sigmoid decomposition.

2. Grades versus the global bypass. With the model frozen, we fit linear and MLP readouts on the eight grades alone, and remove either the explicit grade residual or the global residual FzF_z from the scoring path (Tables E.7 and E.16).

ReadoutSRCCPLCC
Grade-only, linear0.4430.405
Grade-only, MLP0.5370.484
Grade residual removed0.6480.655
Global residual removed0.7520.762
Full model0.8130.812

The eight grades alone carry overall-score information. Removing the grade residual lowers SRCC to 0.648, whereas removing the global residual lowers it only to 0.752: the prediction depends more on the named grades than on this bypass.

3. Prediction interventions. The reference run achieves 0.813/0.812 SRCC/PLCC. Values report mean absolute prediction changes as a percentage of the full score range (Section E.7, Table E.8).

AnalysisInterventionChange (%)
Grade replacementGrades from training images0.82%
Random noise grades0.32%
Uniform grades0.76%
Binary grades0.76%
Term deletionLargest-magnitude term2.97%
Random term0.73%
Largest term after weight mismatch2.94%
Image edits (N = 813)Edit targeting the top-routed rubric2.93%
Edit targeting a random rubric0.95%
Fixed sharpness edit1.21%

Grade replacement changes predictions, while deleting the largest-magnitude term causes approximately four times the change of random deletion. Under the tested edit protocol, targeting the top-routed rubric produces approximately three times the change of random-rubric edits.

4. Text-anchor semantics after textual LoRA. For two paraphrased reference sets, exact cosine nearest-neighbor matches among the 24 anchors change from 19/24 to 15/24 and 21/24 to 19/24 after adaptation (Section E.7). A majority of anchors therefore retrieve their intended paraphrases after adaptation, although retrieval agreement decreases. This embedding-level check complements the grade-level prompt robustness and human-reference evaluations reported under Q1.

W1. Evidence for rubric-grounded interpretability.

Thank you for highlighting this concern. We agree that overall-score accuracy alone does not establish rubric-level interpretability. The additional evaluations in Q1 and Q2 separately examine what the grades represent and how they participate in prediction.

Semantic alignment (Q1). We now evaluate against human annotations, complementing the original exploratory ArtiMuse-F comparison. On EVA, PARA, and the eight-rubric expert-annotated dataset, RLA-Net achieves macro RMSE of 0.172, 0.134, and 0.241, compared with 0.248, 0.314, and 0.271 for AesCLIP-Ens, respectively. These results provide direct human-reference evidence on the evaluated attributes. Dataset-wide distributions, paraphrasing tests, and paired image edits additionally examine grade variability, stability, and differentiated responses.

Predictive role and score composition (Q2). Readouts using only eight grades achieve 0.443 and 0.537 SRCC, demonstrating predictive information without contextual descriptor inputs. Removing the explicit grade residual lowers SRCC from 0.813 to 0.648, compared with 0.752 when the global residual is removed. We verify that the complete contribution terms reconstruct the pre-sigmoid prediction to a maximum error of 2.38×10−72.38\times10^{-7}, and report grade-replacement and term-deletion tests that assess prediction sensitivity.

These results support rubric estimates with human-reference agreement and predictive utility within an explicit score decomposition. They do not establish human-aligned causal attribution: predictions retain contextual features, and routing weights represent internal aggregation rather than human-validated importance. See Q1 and Q2 and the updated supplementary PDF for details.

AuthorsOpenReview ↗

Response to Reviewer zD3u - Part III (W2)

W2. Controlled comparisons, repeatability, and component attribution.

Thank you for highlighting these concerns. We have added matched baseline comparisons, repeated training runs, and separate routing/gating ablations, and clarified the dataset-specific protocols.

1. Matched baseline comparisons. We standardized the comparison setting by evaluating AesCLIP and FGAesQ after training on the corresponding target datasets, using the same data splits and evaluation protocol as RLA-Net. We updated Table 2 of the main paper, replacing the earlier checkpoint-only and transfer entries with these results. The remaining baseline values are literature-reported results.

DatasetMethodSRCCPLCC
PARAFGAesQ0.8990.928
PARARLA-Net0.9270.951
TAD66KAesCLIP0.4820.509
TAD66KFGAesQ0.4750.502
TAD66KRLA-Net0.5080.539
FLICKR-AESAesCLIP0.7870.812
FLICKR-AESFGAesQ0.7890.809
FLICKR-AESRLA-Net0.8120.832

RLA-Net outperforms the evaluated retrained baselines on each dataset. On FLICKR-AES, it exceeds the strongest retrained SRCC and PLCC by 0.023 and 0.020, respectively.

2. Repeated training runs on AVA subset. To assess training variability, we trained three configurations with seeds 1, 2, and 43 on a reduced AVA subset. We report the mean test SRCC and sample standard deviation (Section E.8, Table E.9).

ConfigurationSRCC (mean ± s.d.)
Full model0.813 ± 0.009
Without CRA0.794 ± 0.012
Two-level anchors0.794 ± 0.008

The mean SRCC gains are 0.019 over the no-CRA variant and 0.019 over the two-level variant, both larger than the observed run-to-run standard deviations.

3. Separating routing and gating. We retrained AVA variants that vary routing and gating separately under the same experimental protocol (Section E.3, Table E.3).

RoutingGateSRCCPLCC
Image-conditionedLearned (full model)0.8130.812
UniformLearned0.6480.657
Image-conditionedFixed at 0.50.7120.714
UniformFixed at 0.50.6510.661
Image-conditionedFixed at 00.6590.661
Image-conditionedFixed at 10.6790.689

With the gate fixed at 0.5, image-conditioned routing improves SRCC/PLCC by 0.061/0.053 over uniform routing. With image-conditioned routing, learned gating improves SRCC/PLCC by 0.101/0.098 over the fixed-0.5 gate; both endpoint gates also perform worse than the 0.5 mixture. Learned gating does not improve correlations under uniform routing, indicating that its benefit depends on the routing configuration. These controls distinguish the effects of the two mechanisms and support their combined use in the full model.

4. Dataset-specific protocols. Section E.4, Table E.4 documents the split sources and supervision for each dataset. PARA and TAD66K use their official training and test partitions. On FLICKR-AES, RLA-Net and the retrained baselines use identical fixed data splits and the same evaluation protocol. RLA-Net uses no per-image annotations beyond overall ratings; PARA attribute annotations and TAD66K theme labels are excluded from training. The optimization settings, loss coefficients, and model-selection rule are provided in Section 4.2 and Appendix A.

Together, the matched baseline comparisons, repeated runs on an AVA subset, separate routing/gating ablations, and clarified training protocols strengthen the control and transparency of our evaluation. We hope these additions help address your concerns about comparison fairness, training variability, and component attribution.

Closing: We thank the reviewer for these suggestions, which have strengthened the fairness, repeatability, and component-level evaluation of our approach. Thank you again for your time, effort, and constructive feedback; we welcome any follow-up questions and will respond promptly.

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

The paper reformulates image aesthetic assessment as rubric-grounded score composition under image-level supervision, aiming to jointly recover readable per-rubric states and their image-dependent scoring roles without per-image attribute annotations. On AVA, PARA, TAD66K, and FLICKR-AES, the proposed RLA-Net reports competitive or state-of-the-art SRCC/PLCC.

Critical strengths

The CRA objective is well-motivated and nicely analyzed.

Critical weaknesses

Several key ablation gaps are too small to support the claims drawn from them, e.g., on AVA, (e) 0.813/0.812 vs. (f) no-CRA 0.794/0.793, (h) 2-level 0.794/0.793.

Questions for authors

1 The rubric grades in Figure 3 are concentrated near 0.5 across visually different images. Please analyze their distributions and discriminative ability on a larger test set.

2 Table 3 evaluates rubric routing and gating jointly, making it difficult to determine the individual contribution of each component.

3 Table 2 combines results obtained under different training protocols, particularly on FLICKR-AES, so comparisons with baselines trained under matched conditions would be more convincing.

4 The ablation studies are conducted only on AVA, and additional experiments on other datasets would help establish whether the observed improvements generalize.

1 comment
AuthorsOpenReview ↗

Response to Reviewer ihzx

Response Summary: We thank the reviewer for the careful assessment and constructive comments. We appreciate your positive evaluation of the paper’s clarity and CRA’s motivation and analysis. We are also grateful for your questions, which have helped us clarify our approach and strengthen the paper with additional evidence. We have responded to each weakness and question point by point and conducted additional experiments. We include all new experiments and results in the updated PDFs.

W1&Q4: Ablation effect sizes and evidence beyond AVA.

On AVA, the full model improves over both the no-CRA and two-level-anchor variants from 0.794/0.7930.794/0.793 to 0.813/0.8120.813/0.812 SRCC/PLCC: absolute gains of 0.0190.019 in both metrics, or approximately 2.42.4\\% relative increases. For context, the reported AVA gap between IAACLIP and AesCLIP in Table 2 is 0.017/0.0120.017/0.012. Our ablation gains are therefore comparable in magnitude to reported differences among competitive methods on this benchmark.

We thank the reviewer for suggesting a more complete ablation analysis. We have therefore extended the controlled ablations to PARA, with results included in Section E.1, Table E.1 of the updated supplementary PDF:

PARA configurationSRCCPLCC
Full model (three-level anchors)0.9270.951
Without CRA0.8810.916
Uniform routing0.8090.842
Fixed gate (βc=0.500\beta_c=0.500)0.7830.806
Uniform routing + fixed gate (βc=0.500\beta_c=0.500)0.6730.710
Two-level anchors0.9020.924
Five-level anchors0.8220.865

On PARA, removing CRA and using two-level anchors reduce SRCC/PLCC by 0.046/0.035 and 0.025/0.027, respectively. These additional ablations extend our original AVA-only analysis to PARA, supporting the predictive benefits of directional guidance, three-level semantic references, and content-dependent score composition across both datasets.

Q1: Rubric-grade distributions across the AVA test set.

Thank you for suggesting a dataset-level analysis. We examined the eight rubric grades used by the scoring model on the AVA test images. Table E.2 summarizes their means and medians, while Figure E.1 presents the full histograms. Both are included in Section E.2 of the updated supplementary PDF.

RubricMean gradeMedian grade
Lighting0.450.45
Composition0.480.47
Color0.710.72
Sharpness0.140.12
Contrast0.420.40
Exposure0.340.24
Texture0.500.49
Mood0.480.44

The dataset-level distributions are not uniformly concentrated around a fixed value: color and sharpness have markedly different centers. The histograms in Fig. E.1 also reveal asymmetric distributions, such as exposure, that cannot be characterized by the mean alone. The new analysis provides dataset-level evidence of across-image variation within individual rubrics.

Q2: Separating the contributions of routing and gating.

We added AVA ablations in which routing and gating are varied separately, with each variant retrained under the same experimental protocol. The results are included in Section E.3, Table E.3 of the updated supplementary PDF.

RoutingGateSRCCPLCC
Image-conditionedLearned (full model)0.8130.812
UniformLearned0.6480.657
Image-conditionedFixed at 0.50.50.7120.714
UniformFixed at 0.50.50.6510.661
Image-conditionedFixed at 000.6590.661
Image-conditionedFixed at 110.6790.689

Routing: With the gate fixed at 0.50.5, image-conditioned routing improves SRCC/PLCC by 0.061/0.053 over uniform routing.

Gating: With image-conditioned routing, learned gating improves SRCC/PLCC by 0.102/0.098 over the fixed 0.50.5 gate.

These controls separate the two effects and reveal their interaction: learned gating benefits prediction with image-conditioned routing, but not with uniform routing, supporting their complementary use.

Q3: Consistent target-dataset training for baseline comparisons.

Thanks for the suggestion. We have unified the experimental settings for these comparisons, using the corresponding target training sets and the same data splits and evaluation protocol. On FLICKR-AES, RLA-Net achieves 0.812/0.832 SRCC/PLCC and remains ahead of both target-trained baselines. We have updated Table 2 in the main paper.

MethodDatasetSRCCPLCC
AesCLIPTAD66K0.4820.509
AesCLIPFLICKR-AES0.7870.812
FGAesQPARA0.8990.928
FGAesQFLICKR-AES0.7890.809
FGAesQTAD66K0.4750.502

Closing: Thank you again for your time, effort, and constructive feedback; we welcome any follow-up questions and will respond promptly.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.