{"id":"e6fc29d7-708d-4f2c-8820-cf7608911c4e","arxiv_id":"1908.04347","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Using the Normalised Laplacian Pyramid Distance as a regularizer in conditional GANs yields modest improvements in generated image realism over L1 loss on some datasets.","lead":"A new regularizer for image-to-image GANs replaces the pixel-by-pixel L1 loss with the Normalised Laplacian Pyramid Distance, a metric inspired by the human visual system. The authors report modest improvements in realism metrics on some datasets, but the differences are often within noise.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported gains are too weak statistically to support the conclusion: BRISQUE/NIQE lack error bars, FCN scores overlap, and AMT preference is near chance on two datasets; the paper's own numbers contradict 'better in all evaluation procedures'.","rationale":"The paper's method is plausible and builds on a published perceptual metric, so I am not questioning the internal construction of NLPD. Rather, the most load-bearing weakness is that the reported empirical evidence does not establish the claimed improvement. The reader's verdict is already CONDITIONAL for overlapping standard deviations and near-chance human preference; my analysis points to the same evidence but locates the decisive issue at the level of statistical support rather than at the cross-dataset transfer of the NLPD normalization filters. The manuscript's own Section 4.3 numbers contradict the unqualified Conclusion, and this is independent of any choice about McGill weights. I would keep the verdict CONDITIONAL but require concrete uncertainty quantification as the condition for acceptance.","tokens_in":6274,"tokens_out":4626,"duration_ms":50825,"concrete_test":"Ask the authors for (or recompute from generated outputs) per-image BRISQUE/NIQE scores and the per-decision AMT outcomes, then compute paired 95% bootstrap confidence intervals for the mean differences and exact binomial confidence intervals for the preference proportion for each dataset, across the three random seeds. If any preference interval includes 50% or any quality-difference interval includes 0, the headline claim 'better in all evaluation procedures' should be retracted or replaced by a dataset-specific, uncertainty-quantified statement.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that replacing the L1 regularizer with NLPD makes cGAN outputs perceptually more realistic, as evidenced by FCN segmentation accuracy, BRISQUE/NIQE, and AMT preference. For that claim to hold, the reported improvements have to be distinguishable from noise. They are not, on the paper's own numbers. Table 1 reports BRISQUE/NIQE as point estimates with no variance, so the Facades row (identical 30.1/5.2) and the Cityscapes/Maps differences (2.1 and 1.6 BRISQUE points) could be seed-to-seed noise. Table 2 shows overlapping standard deviations for every FCN metric (per-pixel 0.71±0.15 vs 0.74±0.09; per-class 0.25±0.05 vs 0.25±0.04; IOU 0.18±0.04 vs 0.19±0.04), so the claimed segmentation improvement is not established. The AMT result is 50.04% for Facades and 52.37% for Maps, both compatible with chance given 500 decisions; only Cityscapes (56.16%) is suggestive. The conclusion nevertheless states 'better performance in all the evaluation procedures' and 'users preferred' NLPD images. This is an internal inconsistency in the manuscript's reported numbers, and it is load-bearing because no other quantitative evidence ties NLPD to human-perceptual quality. The NLPD mechanism may still work, but as presented the central claim is statistically under-supported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes replacing the L1 reconstruction regularizer in the pix2pix conditional GAN objective with the Normalised Laplacian Pyramid Distance (NLPD), a perceptual metric inspired by the early human visual system. The authors evaluate this modification on three image-to-image translation datasets (Facades, Cityscapes, Maps) using FCN segmentation accuracy, the no-reference image quality metrics BRISQUE and NIQE, and an Amazon Mechanical Turk preference experiment. They report that the NLPD-regularised model achieves better results than the L1 baseline across these evaluations and conclude that the proposed objective shows better performance in all evaluation procedures.","tokens_in":6613,"tokens_out":3969,"duration_ms":37316,"significance":"If the empirical claim were statistically established, the paper would make a useful contribution: it demonstrates that a biologically motivated, differentiable perceptual distance can serve as a practical training loss for conditional GANs, with a computational overhead that is negligible relative to GAN training. The experimental design covers multiple datasets and standard evaluation protocols, and the motivation for moving beyond pixel-level losses is well argued. However, as presented, the quantitative evidence is not strong enough to support the central claim of consistent perceptual improvement, so the paper's current value lies mainly in the proposal and the experimental setup rather than in the demonstrated gains.","major_comments":[{"comment":"The BRISQUE and NIQE scores are reported as point estimates without any measure of variance or statistical test; the Facades row shows identical scores of 30.1 (5.2) for both methods, and the Cityscapes and Maps differences (2.1 and 1.6 BRISQUE points) could easily arise from seed-to-seed variability. Without error bars or significance tests, the claim that NLPD improves no-reference IQM scores is not established.","section":"Section 4.3, Table 1"},{"comment":"The FCN segmentation metrics show overlapping standard deviations between cGAN+L1 and cGAN+NLPD: per-pixel accuracy 0.71 ± 0.15 versus 0.74 ± 0.09, per-class accuracy 0.25 ± 0.05 versus 0.25 ± 0.04, and class IOU 0.18 ± 0.04 versus 0.19 ± 0.04. Since no significance test or confidence interval is reported, the claimed improvement in segmentation accuracy is statistically indistinguishable from noise.","section":"Section 4.3, Table 2"},{"comment":"The AMT preference results are 50.04% for Facades and 52.37% for Maps, both compatible with chance given 500 decisions per dataset; only Cityscapes (56.16%) is suggestive. The conclusion in Section 5 that 'the suggested objective shows better performance in all the evaluation procedures' and that 'users preferred' the NLPD images is therefore not supported by the presented numbers. This is a load-bearing internal inconsistency because the central claim of perceptual improvement rests on these evaluations.","section":"Section 4.3, AMT; Section 5"},{"comment":"The NLPD normalization filters are precomputed on the McGill dataset from monochromatic images, and the paper provides no analysis of whether these weights remain valid for the target datasets when the NLPD is used as a training loss under GAN optimization. A concrete test would be to compare performance with filters re-estimated on the target data or to validate the transferred metric on human-rated images from the target domain; without this evidence, the risk that the regularizer induces artifacts specific to domain mismatch remains unaddressed.","section":"Section 4.1"}],"minor_comments":[{"comment":"The author byline contains a typo: 'Santos-Rodgriguez' should likely be 'Santos-Rodriguez'.","section":"Author byline"},{"comment":"The phrase 'above chance' should be accompanied by a confidence interval or a binomial test; as written, it is not a statistical claim.","section":"Section 4.3"},{"comment":"The caption 'BRISQUE (NIQE) Scores' is ambiguous; it should clarify that both BRISQUE and NIQE scores are reported.","section":"Table 1 caption"},{"comment":"The qualitative descriptions of texture and contrast differences between methods are not quantified; adding a small-scale quantitative comparison or a separate perceptual study would strengthen these observations.","section":"Section 4.3, Figures 1 and 2"},{"comment":"The paper does not state whether code or trained models will be released, which would aid reproducibility of the proposed regulariser.","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The main barrier to acceptance is the statistical under-support of the central claim; the authors should add significance tests or confidence intervals for the BRISQUE/NIQE and AMT results and soften the conclusion to match the evidence. Also note that one co-author is a developer of the NLPD metric; while this is not inappropriate, the editors may wish to ensure the paper includes a conflict-of-interest statement or an explicit acknowledgement of this relationship."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is the first application of NLPD as a GAN regulariser, replacing L1 in the pix2pix objective. The idea is sensible and the implementation is straightforward, but the paper's own numbers do not support the conclusion that NLPD is better across all evaluations.\n\nWhat is genuinely new is narrow: NLPD was previously an image quality metric and a rendering regulariser; using it as a cGAN training loss is a new combination. The authors use standard datasets and splits, the same architecture as pix2pix, and they honestly report three different evaluation strategies. That is more than many GAN papers do.\n\nThe soft spots are real. In the FCN table every metric overlaps between L1 and NLPD when you account for standard deviation; per-pixel accuracy is 0.71±0.15 against 0.74±0.09. The BRISQUE/NIQE table has point estimates with no variance, so the Cityscapes and Maps differences could be seed noise. The Facades row is identical for both losses, and the AMT preference is 50.04% on Facades, which is chance. Yet the conclusion states 'better performance in all the evaluation procedures' and that users preferred the NLPD images. That internal inconsistency is load-bearing, because without the unqualified conclusion the evidence reduces to 'sometimes better, sometimes indistinguishable.'\n\nI do not think the central idea is wrong. NLPD is psychophysically motivated, its normalisation weights were fit on McGill rather than the target datasets, and it has been externally validated as an IQM, so the self-citation concern is minor. But the authors need to provide error bars or significance tests, and they need to soften the conclusion to match the data. The cross-dataset transfer of the normalisation filters is also assumed rather than tested; a small ablation would help.\n\nWho should read this: people working on perceptual losses for conditional GANs, and anyone teaching evaluation of generative models. I would send it to a competent referee, not desk-reject it; the question is answerable and the method is already public. I would not cite it as evidence of improvement until the statistical issues are addressed.","headline":"A plausible drop-in perceptual loss for cGANs that is weakened by overclaimed conclusions and statistics that do not support them.","tokens_in":7166,"tokens_out":2531,"would_cite":false,"duration_ms":24036,"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":"Replacing the L1 regulariser with the Normalised Laplacian Pyramid Distance (NLPD) in a cGAN objective produces more perceptually realistic generated images, as measured by segmentation accuracy, no-reference quality metrics, and human…","keywords":["conditional generative adversarial networks","image-to-image translation","perceptual image quality assessment","Normalised Laplacian Pyramid Distance","perceptual regulariser","L1 loss","FCN segmentation","no-reference image quality metrics"],"falsifier":"Re-run the Cityscapes training with NLPD normalisation filters recomputed from the target dataset's own statistics rather than taken from the precomputed monochromatic set; if FCN per-pixel accuracy does not rise above the L1 baseline, the transfer of the normalisation filters is the load-bearing component.","tokens_in":6084,"feed_emoji":"🖼️","tokens_out":7005,"duration_ms":68436,"temperature":0.7,"pith_summary":"The paper proposes a way to put human perception into the training objective of conditional generative adversarial networks (cGANs) for image-to-image translation. It replaces the standard L1 reconstruction regulariser with the Normalised Laplacian Pyramid Distance (NLPD), a multi-scale, locally normalised distance derived from the early stages of the human visual system. The claim is that this swap makes generated images more perceptually realistic: on the Cityscapes, Maps, and Facades benchmarks, the NLPD-regularised model improves FCN segmentation accuracy and no-reference image quality scores, and human raters prefer its output over L1-regularised output. If true, the result matters because it shows that a principled perceptual distance can be used directly as a learning signal, rather than being reserved for evaluation.","feed_headline":"NLPD regulariser beats L1 in GAN image realism tests","feed_subtitle":"Swapping L1 for a human-visual-system distance improves segmentation, no-reference scores, and human preference.","key_machinery":"The central object is the Normalised Laplacian Pyramid Distance (NLPD), a perceptual image distance defined in the paper's foundational reference [8] and used here as a cost term. The image is filtered and downsampled through several stages; at each stage each coefficient is divided by a local estimate of the amplitude or energy, mimicking contrast normalisation in the human visual system. The distance between two images is the root-mean-square error between their transformed representations across $N$ stages. In this paper the transform is appended to the cGAN computation graph as extra convolutional layers, so its gradient flows into the generator and enforces realistic local and global contrast during training.","core_discovery":"The paper's central claim is that optimising a cGAN with the objective $L_{cGAN}(G,D) + \\lambda L_{NLPD}$ instead of $L_{cGAN}(G,D) + \\lambda L_{L1}$ yields images that are closer to human perception of natural images. The NLPD term computes the RMS distance between the generator output and the ground-truth image in a 'perceptual space' built from a Laplacian pyramid with local amplitude normalisation at each scale. The authors report that this regulariser produces more realistic local and global contrast, improves per-pixel accuracy and class IOU of an FCN segmentation network trained on ground-truth Cityscapes, and lowers BRISQUE and NIQE scores on Cityscapes and Maps. In a human survey asking which image looks more natural, NLPD-regularised images were preferred above chance for Maps and Cityscapes and at chance for Facades.","pith_inferences":["Editorial inference: the NLPD regulariser is not specific to cGANs; any differentiable image-generation objective with a reconstruction term could be tested with it, including super-resolution or style transfer, although the gradient behaviour of the normalisation filters would need re-checking.","Editorial inference: the authors fix the NLPD normalisation filters from a monochromatic image set and apply them per colour channel; relearning these filters on each target dataset, or jointly with the generator, is a direct testable extension that could improve transfer.","Editorial inference: a natural extension is to test whether combining NLPD with a small L1 term preserves pixel fidelity while adding perceptual quality, since NLPD might behave differently from L1 under strong regularisation."],"forward_implications":["On Cityscapes, FCN per-pixel accuracy rises from $0.71 \\pm 0.15$ with L1 to $0.74 \\pm 0.09$ with NLPD, and class IOU from $0.18 \\pm 0.04$ to $0.19 \\pm 0.04$, suggesting that NLPD-generated images retain more features of the original dataset.","BRISQUE and NIQE scores for the NLPD model move closer to ground-truth values on Cityscapes and Maps, indicating more natural-looking outputs according to no-reference metrics.","Human raters choose NLPD-generated images over L1-generated images in 52.37% of Maps comparisons and 56.16% of Cityscapes comparisons, while Facades is at 50.04%.","The NLPD regulariser's computational overhead is modest relative to GAN training, because the pyramid transform can be expressed as extra convolutional layers in the computation graph.","The results support the broader principle that perceptual metrics validated on human judgment can serve as training objectives, not just evaluation tools."],"supporting_citations":[{"why":"Defines the cGAN baseline, generator/discriminator architecture, L1 regulariser, datasets, and train/test splits that the comparisons reuse.","marker":"[7]"},{"why":"Introduces NLPD and reports its correlation with human perceptual ratings, making it the perceptual distance this paper adopts as a regulariser.","marker":"[8]"},{"why":"Supplies the image set used to precompute the local-amplitude normalisation weights for NLPD.","marker":"[13]"},{"why":"The fully convolutional network used to measure segmentation accuracy on generated Cityscapes images.","marker":"[10]"},{"why":"BRISQUE, one of the two no-reference image quality metrics used to score naturalness of generated images.","marker":"[11]"},{"why":"NIQE, the second no-reference image quality metric used to score generated images.","marker":"[12]"},{"why":"The Cityscapes dataset, the main benchmark for the FCN segmentation evaluation.","marker":"[3]"},{"why":"The Facades dataset used in the image-to-image translation experiments.","marker":"[16]"}],"fun_headline_variants":["NLPD beats L1 for realistic GAN images","Perceptual GAN regulariser improves image realism","Normalised Laplacian Pyramid Distance improves GAN outputs","GANs get more realistic with NLPD regulariser","NLPD regulariser boosts GAN image quality"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the NLPD normalisation filters, precomputed on a monochromatic image set, remain valid for colour target datasets when used as a gradient signal during GAN training; if these filters do not transfer, the regulariser could push the generator toward artifacts instead of perceptual quality.","fun_headline_variants_meta":{"raw":{"variants":["NLPD beats L1 for realistic GAN images","Perceptual GAN regulariser improves image realism","Normalised Laplacian Pyramid Distance improves GAN outputs","GANs get more realistic with NLPD regulariser","NLPD regulariser boosts GAN image quality"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000143,"raw_usage":{"total_tokens":1161,"prompt_tokens":923,"completion_tokens":238,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":539,"completion_tokens_details":{"reasoning_tokens":163}},"tokens_in":539,"tokens_out":238,"duration_ms":2717,"temperature":1.0,"reasoning_tokens":163,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:14:37.168047+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the Cityscapes training with NLPD normalisation filters recomputed from the target dataset's own statistics rather than taken from the precomputed monochromatic set; if FCN per-pixel accuracy does not rise above the L1 baseline, the transfer of the normalisation filters is the load-bearing component.","supporting_citations":[{"cited_title":"Isola, J","cited_arxiv_id":null,"evidence_quote":"Defines the cGAN baseline, generator/discriminator architecture, L1 regulariser, datasets, and train/test splits that the comparisons reuse."},{"cited_title":"Laparra, J","cited_arxiv_id":null,"evidence_quote":"Introduces NLPD and reports its correlation with human perceptual ratings, making it the perceptual distance this paper adopts as a regulariser."},{"cited_title":"Olmos and F","cited_arxiv_id":null,"evidence_quote":"Supplies the image set used to precompute the local-amplitude normalisation weights for NLPD."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The fully convolutional network used to measure segmentation accuracy on generated Cityscapes images."},{"cited_title":"Mittal, A","cited_arxiv_id":null,"evidence_quote":"BRISQUE, one of the two no-reference image quality metrics used to score naturalness of generated images."},{"cited_title":"Mittal, R","cited_arxiv_id":null,"evidence_quote":"NIQE, the second no-reference image quality metric used to score generated images."},{"cited_title":"Cordts, M","cited_arxiv_id":null,"evidence_quote":"The Cityscapes dataset, the main benchmark for the FCN segmentation evaluation."},{"cited_title":"Tyleˇ cek and R.ˇS´ ara","cited_arxiv_id":null,"evidence_quote":"The Facades dataset used in the image-to-image translation experiments."}],"review_version":1}