{"id":"ec1edbf7-aa68-4831-a565-0d3d3d8def78","arxiv_id":"2411.17390","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"DRI-IQA couples contrastive-learned quality and degradation representations with restoration-network guidance and a feature-matching loss to improve no-reference image quality prediction on several standard benchmarks.","lead":"A new no-reference image quality model splits each image's learned representation into separate quality and degradation halves, then uses an image-restoration network to help predict quality scores. It reports small gains over prior work on several benchmarks, but loses to the QPT method on the LIVEC real-world dataset and gives no error bars.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"RS Loss is architecturally disconnected from the IQA scoring path under the 'fixed DRE' description, so the claimed +0.002 SROCC contribution cannot be caused by it.","rationale":"The reader's weakest assumption focused on the small ablation deltas being within run-to-run noise. My concern is stronger: under the described 'fixed DRE' and the fact that the restored image does not enter the score predictor, RS Loss has no gradient path to influence IQA scores at all. This makes the claimed contribution of RS Loss unsupported by the architecture, not merely noisy. I therefore partially agree with the reader. The verdict remains CONDITIONAL because the paper could resolve this by clarifying the gradient path (e.g., if DRE is actually trainable for RS Loss) or by providing a full code release and gradient analysis. If the gradient path is indeed zero and the ablation delta persists, the claim should be retracted, which would move the verdict toward REJECT. The concern is concrete and testable, and it does not rely on disagreement with community consensus or any ad hominem.","tokens_in":12643,"tokens_out":4625,"duration_ms":57234,"concrete_test":"Instrument the model to compute the gradient of L_RS with respect to the quality-representation encoder and the score predictor parameters. If the gradient norm is exactly zero (as implied by the fixed DRE), then rerun the Proposed vs V3 ablation 10 times with different seeds and report mean±std. If the SROCC difference persists despite zero gradient to the scoring path, it is noise. Alternatively, train V3 with NAFNet weights frozen (only DRE and predictor trained) and compare to Proposed; if the difference persists, RS Loss cannot be the cause.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim that RS Loss contributes to IQA gains is not supported by the architecture as described. In §3.4, L_RS = ||U pRref − U pRres||2 is computed from DRE outputs on restored and reference images, and the text explicitly says this uses the 'fixed DRE' (§3.1, Figure 2). The score predictor receives only the low-quality image and its quality representation; the restored image from NAFNet is not fed to the predictor. With DRE frozen, gradients from L_RS update only NAFNet. NAFNet is not in the scoring path, so changes to NAFNet cannot alter predicted MOS. The Table 3 improvement of +0.002 SROCC (Proposed vs V3) must therefore be attributed either to run-to-run noise (the paper reports no standard deviations) or to an unstated gradient path. If DRE is actually trainable in Stage 2, the phrase 'fixed DRE' is misleading and the paper must describe how RS Loss updates the quality representation. This is more fundamental than the small effect size: the proposed mechanism is structurally absent under the stated design.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DRI-IQA, a no-reference image quality assessment method built around a Dual-Representation Extractor (DRE) that splits one encoder's feature vector into a degradation-aware half and a quality-aware half using contrastive learning. In a second stage, a NAFNet restoration network is used as an auxiliary module, and a Representation-based Semantic Loss (RS Loss) is defined between the degradation representations of restored and reference images. Experiments report SROCC and PLCC on LIVE, KADID-10K, KonIQ-10k, LIVEC, and PIPAL, with the paper claiming favorable performance against state-of-the-art on both synthetic and real-world datasets. The evaluation uses external human-MOS benchmarks, so the primary performance comparison is not circular.","tokens_in":13065,"tokens_out":7415,"duration_ms":71842,"significance":"If the claimed results hold, the paper offers a useful contribution: a compact contrastively pretrained dual-representation encoder that avoids large-scale ImageNet pretraining, an interpretable separation of degradation and quality information, and a restoration-assisted scoring pipeline. Table 2 shows strong results on LIVE, KADID-10K, and KonIQ-10k, and the use of external MOS benchmarks makes the core comparison non-circular. The significance is tempered, however, by the LIVEC result where the method trails QPT, by the lack of statistical support for the two headline components in Table 3, and by an under-specified gradient path for RS Loss. The paper also promises code release, which would help reproducibility, but the code is not currently available.","major_comments":[{"comment":"The abstract and §4.5 claim the method 'performs favorably against existing state-of-the-art models on both synthetic and real-world datasets,' but on LIVEC, QPT reports SROCC 0.895 / PLCC 0.914 while DRI-IQA reports 0.861 / 0.886. This is a large gap and it is not discussed anywhere in the paper; it directly undermines the real-world portion of the central claim. Please report this result and either temper the claim or provide an analysis of why DRI-IQA lags on LIVEC.","section":"§4.5, Table 2"},{"comment":"The evidence that restoration assistance and RS Loss improve IQA accuracy is not statistically supported. In Table 3, V3 vs V2 changes PIPAL SROCC/PLCC by +0.003/+0.003 and KonIQ-10k SROCC by +0.000; Proposed vs V3 changes PIPAL SROCC/PLCC by +0.002/+0.007 and KonIQ-10k SROCC/PLCC by +0.002/+0.000. The paper states that results are averaged over 5 seeds (§4.2) but gives no standard deviations or significance tests, so these differences cannot be distinguished from training noise. Please report per-seed results with standard deviations and a significance test, or explicitly weaken the conclusions about the contributions of RAM and RS Loss.","section":"§4.6, Table 3"},{"comment":"The mechanism by which RS Loss improves the score predictor is not established by the text as written. Equation (6) is computed between degradation representations of the restored and reference images using the 'fixed DRE' described in §3.1 and Figure 2, so under that description gradients from L_RS update only NAFNet, and NAFNet's output is not fed to the MOS predictor. If the main-path DRE is actually trainable in Stage 2, the phrase 'fixed DRE' is misleading and the paper must describe the gradient path through which RS Loss affects the quality representation; if the DRE is truly fixed, the +0.002 SROCC gain in Table 3 cannot be caused by RS Loss and should be attributed to noise or to an unstated mechanism. Please clarify whether the DRE used in RSB is a frozen copy or the same trainable encoder, and show explicitly how L_RS influences the scoring path.","section":"§3.4, Eqs. (4)-(6), Figure 2"}],"minor_comments":[{"comment":"The numerical gains reported for KonIQ-10k are inconsistent with Table 2: the text says DRI-IQA gains 0.004, 0.017, and 0.025 SROCC over QPT, Re-IQA, and CONTRIQUE, but the table shows 0.009, 0.022, and 0.042. The text also contains the typo 'KoiIQ-10k'.","section":"§4.5"},{"comment":"The text says V2 gains 0.03 SROCC and 0.06 PLCC on PIPAL compared to V1, but Table 3 shows 0.003 and 0.006. Please correct the decimal values and ensure the prose matches the table.","section":"§4.6"},{"comment":"The experimental protocol says results are 'averaged by 10 times split' and averaged over 5 seeds, but the split procedure is not fully specified. Please describe the train/test split protocol, report standard deviations for Table 2 as well as Table 3, and state which results were copied from prior work versus obtained under this protocol.","section":"§4.2, Table 3"},{"comment":"Since the score predictor in DRI-IQA is based on MANIQA [50], including MANIQA in Table 2 would clarify the gain attributable to the DRI, RAM, and RS Loss modules over the backbone predictor. Currently MANIQA is discussed in the text but absent from the comparison table.","section":"§3.3, Related Work"},{"comment":"Reference [32] is listed as TIP2002 but BRISQUE was published in TIP 2012, and reference [33] duplicates the same paper under a different title entry. Please correct the citation metadata and remove the duplicate. Also, 'Flicker2K' in §4.1 should be 'Flickr2K'.","section":"References and Table 2"}],"recommendation":"major_revision","confidential_remarks":"The main empirical strengths are on LIVE, KADID-10K, and KonIQ-10k, and the evaluation is not circular. The reviewer-level concerns are the unaddressed LIVEC gap, the statistically unsupported ablation deltas, and the ambiguous gradient path for RS Loss. These are fixable within the scope of a revision, so I do not recommend rejection; however, the overclaim in the abstract and §4.5 should be corrected regardless of the other revisions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take on Yue et al. (arXiv:2411.17390). The paper combines known pieces—contrastive dual-representation learning, a NAFNet restoration auxiliary task, and a feature-matching loss—into a NR-IQA model. The main results on LIVE, KADID-10k and KonIQ-10k are slightly better than recent strong baselines (e.g., +0.004 SROCC over QPT on KonIQ, +0.082 over TRes on KADID). That is real, if incremental. The dual-representation split (upper half for degradation, lower for quality) is a neat way to force one encoder to learn both, and the pre-training on only ~3.4k images without ImageNet is a practical plus.\n\nThe soft spots are real but not fatal. First, the paper claims “favorably against state-of-the-art” while Table 2 shows DRI-IQA underperforms QPT on LIVEC (0.861 vs 0.895 SROCC). That needs to be acknowledged and discussed. Second, there are no error bars anywhere, even though all results are averaged over 5 seeds; the main ablation deltas (V3 vs V2: +0.003 SROCC on PIPAL; Proposed vs V3: +0.002) are smaller than typical run-to-run noise. As is, the ablation does not demonstrate that restoration assistance or RS Loss contribute anything beyond noise.\n\nOn the stress-test concern about RS Loss: the worry is that with a frozen DRE, L_RS only updates NAFNet, which is not in the scoring path. On careful reading, the architecture does allow an indirect path: NAFNet is conditioned on reserved features from the (trainable) DRE, so gradients from L_RS can flow back through NAFNet into the DRE and thereby affect the quality representation. But the paper does not spell this out, and calling the DRE “fixed” in §3.4 is misleading. More importantly, the reported +0.002 SROCC contribution is within noise either way. This is an exposition and evidence problem, not a contradiction in the method.\n\nOverall, this is a solid workshop-to-conference-level incremental paper. It is not a breakthrough, but it is a legitimate extension with positive results on standard benchmarks. A referee should ask for error bars, a LIVEC comparison with QPT, a clearer description of the RS Loss gradient path, and an honest discussion of the tiny ablation gains. I would send it to review; with revisions it could be a decent conference paper. I would not cite it in my own work unless I needed a recent NR-IQA baseline.\n\nRecommendation: send to peer review, major revision.","headline":"Reasonable incremental NR-IQA paper; the dual-representation idea is sensible and the synthetic-dataset results are good, but the restoration-assistance contributions are supported mostly by sub-noise ablation deltas and the RS Loss mechanism is under-specified.","tokens_in":13447,"tokens_out":2487,"would_cite":false,"duration_ms":23100,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Splitting a single encoder's features into degradation and quality halves lets a no-reference IQA model outrank prior state-of-the-art on synthetic and real-world benchmarks.","keywords":["No-Reference Image Quality Assessment","Dual-Representation Interaction","Degradation Representation","Quality Representation","Restoration Assistance","Representation-based Semantic Loss","Contrastive Learning","NAFNet"],"falsifier":"Retrain the V2, V3, and Proposed variants (dual-representation only, plus NAFNet, plus RS Loss) on PIPAL and KonIQ-10k with, say, ten seeds each and report per-seed SROCC/PLCC distributions; if the mean deltas fall inside one standard deviation of each other, the paper's component-wise attribution collapses.","tokens_in":12462,"feed_emoji":"🖼️","tokens_out":4366,"duration_ms":37118,"temperature":0.7,"pith_summary":"DRI-IQA is a no-reference image quality assessment method claiming that a distorted image should be represented by two complementary signals—what kind of degradation it has and how bad that degradation looks—rather than by a single quality feature. The paper argues that splitting one encoder's latent representation in half, training the two halves with contrastive learning, and then feeding the degradation half into a restoration network as guidance gives an IQA model the information it needs to score both synthetic and real-world distortions and restored images. If the claim holds, IQA models can be made more robust to real-world distortions and to outputs of restoration models without needing reference images or large-scale pretraining. The paper reports the top SROCC on LIVE (0.982), KADID-10K (0.941), and KonIQ-10k (0.936) among the compared methods.","feed_headline":"One encoder splits quality from degradation to beat IQA benchmarks","feed_subtitle":"DRI-IQA tops LIVE, KADID-10K, and KonIQ-10k correlation tables with restoration-assisted dual representations.","key_machinery":"The Dual-Representation Extractor (DRE): a single five-layer hybrid encoder whose output is split into an upper degradation-aware half and a lower quality-aware half, trained with InfoNCE contrastive loss; the Restoration Assistance Module (RAM), which feeds the degradation representation into NAFNet to restore the image and thereby forces the degradation half to encode restoration-relevant information; and the Representation-based Semantic Loss (RS Loss), an L2 penalty between the degradation representations of the restored image and the clean reference. The split representation is the load-bearing object: it lets one forward pass carry both kinds of information, and the paper argues the two halves interact implicitly through shared layers and losses.","core_discovery":"The central claim, stated on the paper's own terms, is that dual-representation interaction makes a no-reference IQA model perform favorably against existing state-of-the-art models on both synthetic and real-world datasets, and that restoration assistance plus a representation-based semantic loss are what make the degradation information useful to the score predictor. The mechanism: one hybrid encoder produces a shared representation whose upper half is trained to be degradation-aware and whose lower half is quality-aware, using InfoNCE contrastive pairs; in a second stage, the degradation half guides NAFNet through a Restoration Assistance Module, and an RS Loss pulls the degradation representation of restored images toward that of the clean reference, implicitly sharpening the quality representation. On the evaluated benchmarks the model reports the best SROCC/PLCC pairs on LIVE, KADID-10K, and KonIQ-10k, and its ablation attributes the gains to the dual-representation split, the restoration module, and the RS Loss in turn.","pith_inferences":["The dual-representation split is a generic inductive bias: the same half-degradation/half-quality design could transfer to other image-understanding tasks where content and nuisance factors need separate handling, such as face recognition under noise or medical image screening.","The reported gains from RAM and RS Loss are small in absolute terms (at most 0.007 in correlation); unless variance over seeds is tiny, those specific components may not generalize beyond the reported datasets even if the dual-representation split itself carries most of the benefit.","A testable extension would be to swap NAFNet for a GAN-based restoration network—which the paper lists as future work—to probe whether RS Loss's degradation-pulling effect remains useful on restorations that are perceptually strong but structurally different, exactly where the paper notes GAN artifacts fool quality predictors.","Because DRE is trained on high-quality natural images only, the method likely inherits a content bias; cross-dataset transfer tests (e.g., train on KonIQ-10k, test on LIVEC) would show whether the dual representation captures dataset-agnostic degradation factors."],"forward_implications":["If the central claim holds, NR-IQA methods no longer need to choose between quality-aware and degradation-aware features; a single encoder can supply both.","The restoration-assistance setup implies that any restoration network with an encoder-decoder structure can be plugged in to supply degradation guidance, so the approach is not tied to NAFNet itself.","Because the DRE is pretrained only on a small hybrid-degraded set drawn from DIV2K and Flickr2K, the results imply that strong IQA performance does not require ImageNet-scale pretraining.","The RS Loss, by penalizing degradation-representation differences between restored and clean images, gives a training signal that directly targets the artifacts restoration models introduce—an issue the paper identifies in GAN-based restorers."],"supporting_citations":[{"why":"NAFNet is the restoration network inside the Restoration Assistance Module; its guidance is the component whose ablation isolates the benefit of restoration assistance.","marker":"[1]"},{"why":"MoCo-style InfoNCE loss is used to pretrain the Dual-Representation Extractor and to constrain it in both training stages.","marker":"[14]"},{"why":"MANIQA is the Transformer-based IQA network used as the score predictor backbone and the guidance architecture that consumes the quality representation.","marker":"[50]"},{"why":"QPT is the conventional single-quality-representation pipeline that DRI extends and a primary baseline for the dual-representation comparison.","marker":"[57]"},{"why":"The PIPAL dataset is used in the ablation study to demonstrate the incremental effectiveness of the dual representation, restoration assistance, and RS Loss.","marker":"[12]"},{"why":"The LIVE dataset supplies the paper's strongest reported SROCC (0.982) and anchors its synthetic-distortion evaluation.","marker":"[41]"}],"fun_headline_variants":["Split quality and degradation to top IQA leaderboards","Restoration-assisted dual reps beat IQA benchmarks","DRI-IQA: dual encoding with restoration aids IQA","Quality-degradation split improves no-reference IQA"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim stands on the assumption that the averaged correlation gains from adding restoration assistance and the RS Loss—deltas of roughly 0.001 to 0.007—are real improvements rather than run-to-run noise, since the paper reports five-seed averages without standard deviations.","fun_headline_variants_meta":{"raw":{"variants":["Split quality and degradation to top IQA leaderboards","Restoration-assisted dual reps beat IQA benchmarks","DRI-IQA: dual encoding with restoration aids IQA","Quality-degradation split improves no-reference IQA"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000143,"raw_usage":{"total_tokens":1148,"prompt_tokens":896,"completion_tokens":252,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":512,"completion_tokens_details":{"reasoning_tokens":188}},"tokens_in":512,"tokens_out":252,"duration_ms":4540,"temperature":1.0,"reasoning_tokens":188,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:08:58.202912+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the V2, V3, and Proposed variants (dual-representation only, plus NAFNet, plus RS Loss) on PIPAL and KonIQ-10k with, say, ten seeds each and report per-seed SROCC/PLCC distributions; if the mean deltas fall inside one standard deviation of each other, the paper's component-wise attribution collapses.","supporting_citations":[{"cited_title":"Maniqa: Multi-dimension attention network for no-reference image quality assessment","cited_arxiv_id":null,"evidence_quote":"MANIQA is the Transformer-based IQA network used as the score predictor backbone and the guidance architecture that consumes the quality representation."},{"cited_title":"Quality-aware pre-trained models for blind image quality as- sessment","cited_arxiv_id":null,"evidence_quote":"QPT is the conventional single-quality-representation pipeline that DRI extends and a primary baseline for the dual-representation comparison."},{"cited_title":"Pipal: a large-scale image quality assessment dataset for perceptual image restoration","cited_arxiv_id":null,"evidence_quote":"The PIPAL dataset is used in the ablation study to demonstrate the incremental effectiveness of the dual representation, restoration assistance, and RS Loss."},{"cited_title":"A statistical evaluation of recent full reference image quality assessment algorithms","cited_arxiv_id":null,"evidence_quote":"The LIVE dataset supplies the paper's strongest reported SROCC (0.982) and anchors its synthetic-distortion evaluation."}],"review_version":1}