{"id":"f368edd0-79d4-45bd-80ef-2330091aff13","arxiv_id":"2501.14264","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"CDI is a wavelet-domain fidelity metric for blind image restoration that compares restored images with the degraded input rather than the reference, plus a reference-free variant and a new subjective dataset.","lead":"Blind image restoration algorithms often produce visually appealing but unfaithful results, and standard quality metrics punish them for inventing details. This paper proposes CDI, a metric that instead checks how consistently a restored image re-degrades back to the original degraded input, and builds a dataset to test it.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"DISDCD's ground truth asks raters for the same quantity RGCDI computes, so the empirical claim is partly self-fulfilling; independent fidelity-preference validation is needed before 'significantly superior' is accepted.","rationale":"The reader's weakest assumption is exactly the load-bearing concern: DISDCD defines subjective fidelity in the same terms the metric computes. I agree with that reading. The paper has genuine strengths: the two-step degradation model, the noise-equivalent attenuation derivation, and the wavelet-domain formulation are plausible and internally consistent, and the authors are careful to state that CDI only evaluates low-frequency fidelity. However, the empirical validation on DISDCD cannot by itself establish that CDI is the right notion of BIR fidelity, because the raters are asked to compare re-degraded restored images to the degraded image, which is the same criterion RGCDI approximates. The 'significantly superior' claim is therefore conditional on accepting that operationalization. Since the reader already assigned CONDITIONAL with moderate confidence, and no additional internal inconsistency emerged in my pass, the verdict should remain unchanged. The concrete test of an independent reference-based human fidelity preference study would either confirm the claim or reveal that CDI's advantage is an artifact of the dataset design.","tokens_in":17564,"tokens_out":8699,"duration_ms":89822,"concrete_test":"Build a fresh subjective dataset from the same 11 BIR algorithms and 4 degradation types, but do not show re-degraded restored images. Instead, show raters the original reference image and two restored images, and ask which restored image is more faithful to the reference (or rate restoration fidelity on a continuous scale). Compute RGCDI and RACDI agreement with these judgments and compare against PSNR, SSIM, LPIPS, and DISTS. If CDI's advantage shrinks or reverses, the DISDCD result is an artifact of asking raters to perform the CDI computation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on DISDCD, but DISDCD's subjective task is constructed to be the very quantity RGCDI computes. In Sec. 7, raters see a degraded image and two re-degraded restored images (same degradation, fixed noise seed) and are asked which re-degraded image is closer to the degraded image. RGCDI, by design (Sec. 5, Eq. 5), compares adaptively attenuated restored images against a noise-removed, noise-equivalently attenuated version of the degraded image. Thus high agreement in Fig. 11 largely verifies that RGCDI approximates its own target, not that CDI is the correct operationalization of BIR fidelity. The paper redefines fidelity as low-frequency consistency with the degraded image (Sec. 1.2) and then uses a dataset that asks raters to make exactly that judgment. If human fidelity perception also weights high-frequency plausibility, semantic correctness, or naturalness, the DISDCD result is partly self-fulfilling. The problem is compounded by lambda being tuned on DISDCD (Fig. 12), RACDI being validated only against RGCDI pseudo-labels (Table 1), and code/data being unavailable. The analytic core (Eq. 4c/4d, idempotency/cascade properties) is internally coherent; the gap is external validity of the benchmark.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a new fidelity metric for blind image restoration (BIR) that compares a restored image with the degraded image rather than with the reference image. The method, called Consistency with Degraded Image (CDI), is implemented in two variants: RGCDI, a wavelet-domain algorithm that removes noise from the degraded image with reference guidance and applies information-equivalent attenuation, and RACDI, a reference-agnostic variant that approximates RGCDI using a trained Attenuation Net. The authors construct a new subjective dataset, DISDCD, based on pairwise (2AFC) judgments of which re-degraded restored image is closer to the degraded image, and report that RGCDI and RACDI achieve higher agreement with these judgments than existing FR-IQA and NR-IQA methods.","tokens_in":17894,"tokens_out":9137,"duration_ms":77442,"significance":"The core idea of anchoring fidelity evaluation on the degraded image is practically appealing for BIR, where a reference is unavailable. The information-theoretic derivation in Sec. 4, in particular the equivalence between additive noise and equivalent signal attenuation (Eq. 4c = 4d), is a clean theoretical contribution under the stated Gaussian assumptions, and the proposed algorithm has attractive invariance properties if they are correctly proven. The DISDCD dataset is a potentially useful resource for the community, although its design currently conflates the metric's target with the ground-truth protocol. If the authors can validate CDI against an independent measure of human fidelity judgment, the approach would be a valuable complement to existing IQA methods for BIR.","major_comments":[{"comment":"The DISDCD subjective task asks raters to choose which of two re-degraded restored images is closer to the degraded image. This is precisely the quantity that RGCDI computes: RGCDI adaptively attenuates the restored image (Eq. 5) and compares it with the noise-removed, equivalently attenuated degraded image (Sec. 5). Consequently, the high 2AFC agreement in Fig. 11 largely verifies that RGCDI approximates its own target, not that the CDI construct is the correct operationalization of BIR fidelity. The paper must validate CDI against an independent subjective protocol, for example a preference task where raters rank restored images by fidelity to a reference (when available) or by overall perceptual quality, and demonstrate that CDI-predicted rankings correlate with those judgments.","section":"Sec. 7 and Sec. 8.1, Fig. 11"},{"comment":"The hyperparameter λ is tuned by maximizing 2AFC agreement on DISDCD (Fig. 12), and the recommended value of 0.3 is then used for the reported comparisons. Since DISDCD is the same benchmark on which the metric is evaluated, this tuning does not provide independent evidence of generalization. Similarly, RACDI is trained with RGCDI-generated pseudo-labels (Sec. 6) and evaluated only against RGCDI (Table 1), so the reported mean errors measure approximation of RGCDI rather than alignment with human fidelity judgments. The claim of 'significantly superior' would be supported by evaluation on at least one external dataset or a held-out portion of DISDCD using a distinct subjective protocol.","section":"Sec. 5.3, Fig. 12 and Sec. 6, Table 1"},{"comment":"Fig. 11 reports point estimates of 2AFC accuracy without error bars, confidence intervals, or any significance test. The number of raters in DISDCD is not stated, and no inter-rater agreement statistics are provided. The observation that RGCDI scores 'even exceed human scores' is not discussed; a deterministic metric will often exceed the human consistency measure of Eq. 9, which is the expected agreement between two raters, and this should be interpreted carefully. A proper statistical comparison with confidence intervals and a test of differences between methods is needed before the word 'significantly' is used.","section":"Sec. 8.1, Fig. 11"},{"comment":"The paper explicitly redefines BIR fidelity as low-frequency consistency with the degraded image and states that CDI does not evaluate high-frequency signals, which are left to NR-IQA. This is a legitimate design choice, but it changes the meaning of 'fidelity' relative to the standard FR-IQA notion. The conclusion that CDI is 'markedly superior' to FR-IQA methods is then partly a consequence of measuring a different construct. The paper should prominently and consistently frame CDI as a new, complementary evaluation axis rather than as a general replacement for fidelity assessment, and the experimental comparisons should be interpreted in that light.","section":"Sec. 1.2 and Sec. 9"}],"minor_comments":[{"comment":"There are several typos and inconsistent terms: 'In stead' in the abstract, 'Gussian' in the Fig. 4 caption, 'HSV' should be 'HVS' in Sec. 4.2, and 'Noise Equivalent Attention' should be 'Noise Equivalent Attenuation' in Sec. 6.","section":"Abstract, Sec. 4.2, Sec. 6, Fig. 4"},{"comment":"The DISDCD dataset description lacks important experimental details: the number of raters, rater screening and training, the number of trials per rater, and the exact meaning of '5 pairwise judgments for each sample' need to be specified so that the reliability of the subjective ground truth can be assessed.","section":"Sec. 7"},{"comment":"The comparison with NR-IQA methods (NIQE, MANIQA) is not a like-for-like comparison because these methods measure perceptual naturalness, not fidelity. The statement that they 'cannot evaluate fidelity' is true, but the figure and text should acknowledge that the comparison spans different constructs.","section":"Sec. 8.1"},{"comment":"The proof of idempotency for the case F(F(x,y), y) contains an algebraic error: when computing μ'_A F(x,y) = COV(y, F(x,y)) / COV(F(x,y), F(x,y)), the factor α = μ_A / sqrt(1 + σ_D²/σ_H²) is omitted, leading to the incorrect intermediate assertion μ'_A F(x,y) = μ_A x. The final equality F(F(x,y), y) = F(x,y) holds only after the correction, so the derivation as written should be revised.","section":"Supplementary Material, Sec. 12, Eq. 17"},{"comment":"The paper states that the source code and DISDCD dataset will be publicly available 'shortly'. For reproducibility, please provide a URL or a clear availability statement, and include the dataset creation details (e.g., degradation parameters, seed handling, and image sources) in the supplementary material.","section":"Abstract and Sec. 7"},{"comment":"The reported mean errors between RACDI and RGCDI lack standard deviations and sample sizes. Adding these statistics would help the reader judge the consistency of the approximation across different tasks and images.","section":"Table 1 and Tables 8-10"}],"recommendation":"major_revision","confidential_remarks":"The circularity between the proposed metric and the DISDCD subjective protocol is the central weakness. The paper would be stronger if the authors positioned CDI as a specific operational definition of 'consistency with the degraded image' rather than as 'fidelity' in a general sense, and if they provided an independent validation of the metric's perceptual relevance. Given that the analytic core is sound and the dataset could be a useful community resource, a major revision is appropriate rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe headline is that this paper has a real idea but overclaims its empirical support. The CDI concept—score a blind restoration by re-degrading it and comparing to the input—is a sensible reframing of BIR fidelity, and the authors build a clean analytic story around it: wavelet-domain noise splitting, noise-equivalent attenuation under a Gaussian/HVS model, and an adaptive attenuation step that avoids requiring exact degradation parameters. The idempotency and cascade properties are checked and the math is internally consistent. The reference-agnostic variant, trained with a denoising net to predict the attenuated reference from the degraded image alone, is a useful extension, and Table 1 shows reasonable error against RGCDI. The paper is also honest about what CDI covers: it explicitly says it only evaluates low-frequency fidelity and that NR-IQA remains complementary.\n\nThe soft spot is exactly where the stress-test note lands. DISDCD's subjective task asks raters to pick which re-degraded restored image is closer to the degraded image. That is the same quantity RGCDI computes. So the high 2AFC agreement in Fig. 11 partially verifies that RGCDI approximates its own target, not that CDI is the correct operationalization of BIR fidelity. The problem is compounded by lambda being tuned on the same dataset and RACDI being compared only to RGCDI pseudo-labels. There are no error bars or inter-rater statistics for Fig. 11, and code and data are still promised rather than available.\n\nThat said, I would not call the central claim wrong. The paper makes a principled case that, in the ill-posed BIR setting, there is no ground truth reference to compare against, and the only observable constraint is consistency with the degraded observation. That is a legitimate definitional stance. But the paper presents it as an empirical finding of superiority, not as a definition, and that is where the evidence needs to catch up. A fair revision would need either an independent validation set built on a different definition of fidelity (e.g., preference for semantic correctness, naturalness, or downstream task utility) or a transparent argument that closeness to the degraded image is the only defensible target for BIR fidelity.\n\nWho is this for? Researchers working on BIR evaluation and IQA will want to read it; the CDI framework is a useful addition to the toolbox even if the present validation is not conclusive. It deserves a serious referee, but I would expect significant revisions before acceptance. If I were reviewing, my recommendation would be major revision with the dataset question front and center.","headline":"A promising but self-referential metric: the paper defines fidelity as consistency with the degraded image and then validates on a dataset that asks raters for exactly that judgment.","tokens_in":18403,"tokens_out":2326,"would_cite":true,"duration_ms":20649,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68U10","94A08"],"pacs":[],"model":"deepseek-v4-flash","headline":"A restored image's fidelity should be judged by how well it re-explains the degraded photo, not by similarity to a pristine reference; a wavelet consistency score does this without knowing the degradation.","keywords":["blind image restoration","image quality assessment","fidelity evaluation","consistency with degraded image","wavelet domain","reference-agnostic metric","degraded-image comparison dataset","generative restoration quality"],"falsifier":"A decisive test would be a preference study where raters choose which restoration is more faithful to the original scene, with the pristine original shown as reference, among restorations that have identical low-frequency consistency with the degraded image but different high-frequency realism; if raters consistently prefer one, CDI misses part of fidelity. A simpler check is to compute CDI on a dataset whose ground truth was collected against the pristine original rather than by re-degrading restorations.","tokens_in":17380,"feed_emoji":"🖼️","tokens_out":8869,"duration_ms":73325,"temperature":0.7,"pith_summary":"Blind image restoration is ill-posed: from one degraded photo, many visually different restorations are all valid, so the pristine reference is only one solution and cannot be the ground truth for fidelity. The paper argues that fidelity should instead be measured as consistency with the degraded image (CDI), since degradation destroys high-frequency information irreversibly and only the low-frequency content can be checked. It introduces a wavelet-domain Reference Guided CDI (RGCDI) that removes noise from the degraded image, converts noise into an information-equivalent attenuation, and adaptively attenuates the restoration before computing PSNR. A reference-agnostic version (RACDI) learns the same attenuation directly from the degraded image, removing the need for a reference in real-world use. On a new subjective dataset (DISDCD), the CDI scores agree with human fidelity judgments far better than PSNR, SSIM, LPIPS, and DISTS.","feed_headline":"Score blind restoration by the degraded photo, not the original","feed_subtitle":"A wavelet consistency metric beats full-reference IQA on human fidelity judgments and works without a reference.","key_machinery":"The load-bearing mechanism is the wavelet-domain consistency transform $F(x,y)=\\mu_A x/\\sqrt{1+\\sigma_D^2/\\sigma_H^2}$, built on a two-step degradation model $y=\\mu_A x+n_D$ for wavelet coefficients and an HVS mutual-information model. Reference-guided noise removal splits the degraded image into an attenuated signal and independent noise, the noise-equivalent attenuation step turns noise-induced information loss into an equivalent extra attenuation, and adaptive attenuation picks the least-squares coefficient $\\mu_M$ that best fits the restoration to the attenuated reference before computing PSNR. RGCDI's idempotency and cascade-degradation properties make the transform stable under repeated and combined degradations. The reference-agnostic version replaces the reference-guided noise removal with a denoising backbone trained to predict $F(x,y)$ from the degraded image alone, so consistency becomes computable without a reference.","core_discovery":"The paper's central claim is that full-reference fidelity evaluation is the wrong yardstick for blind image restoration. Because the original image is lost during degradation, any restoration that re-degrades to the observed degraded image is a legitimate solution, and high-frequency details generated by restoration are unverifiable in principle. The paper therefore defines fidelity as the consistency between the restored image and the degraded image in low-frequency content, and constructs RGCDI to compute that consistency without knowing the degradation parameters: it decomposes the degraded wavelet coefficients into attenuation plus noise, converts the noise into equivalent attenuation through an HVS mutual-information model, adaptively attenuates the restoration to best match the attenuated reference, and reports PSNR in the image domain. RACDI replaces the reference-guided part with a trained attenuation network, so the same consistency score is available without any reference image. The paper's experiments on DISDCD show these scores track human fidelity choices substantially better than PSNR, SSIM, LPIPS, and DISTS, and also better than running those full-reference metrics on the re-degraded image.","pith_inferences":["Editorial inference: DISDCD asks raters which re-degraded restoration is closer to the degraded image, which is exactly what CDI measures; a fidelity benchmark where raters compare against the pristine original could give reference-based metrics more credit than this protocol does.","Editorial inference: The paper does not test combining CDI with NR-IQA, though its orthogonality claim suggests a two-axis evaluation (fidelity versus perceptual quality) rather than a single score.","Editorial inference: The proven inequality RGCDI PSNR >= PSNR means CDI values are on a different scale from classical PSNR; adopters should compare CDI differences within one evaluation rather than against published PSNR tables.","Editorial inference: If CDI is used as a training loss, it will reward restorations that re-degrade to the input and ignore the realism of invented detail, an interaction with the perception-distortion tradeoff the paper does not discuss."],"forward_implications":["RGCDI and RACDI reach higher 2AFC agreement with human fidelity judgments on DISDCD than PSNR, SSIM, LPIPS, DISTS, and FR-IQA computed on re-degraded images.","Full-reference scores demonstrably reward blurring a restoration, so evaluation pipelines built on them will push blind restoration methods toward smoother, less faithful outputs.","RACDI's mean error relative to RGCDI stays below roughly 2.7 dB across deblurring, super-resolution, denoising, and blind restoration tasks, making reference-free fidelity evaluation practical.","Because CDI deliberately ignores generated high-frequency content, the paper positions it as orthogonal to no-reference quality assessment; a complete BIR evaluation would supplement a CDI fidelity score with an NR-IQA perceptual quality score."],"supporting_citations":[{"why":"Supplies the wavelet-domain two-step degradation model and HVS mutual-information formulas used for noise removal and noise-equivalent attenuation.","marker":"[32]"},{"why":"Provides the Real-ESRGAN second-order degradation pipeline used to generate combined degradation and DISDCD test images.","marker":"[37]"},{"why":"DIV2K is the source image pool for DISDCD and for the blur-sensitivity demonstration that motivates CDI.","marker":"[2]"},{"why":"LPIPS is a perceptual baseline and the 2AFC protocol for scoring DISDCD and computing human accuracy.","marker":"[58]"},{"why":"SSIM is the main full-reference baseline whose scores increase under added blur, motivating the shift to CDI.","marker":"[39]"},{"why":"DISTS is the second perceptual full-reference baseline compared on DISDCD.","marker":"[10]"},{"why":"Supplies the SCUNet backbone trained as the Attenuation Net for the reference-agnostic RACDI.","marker":"[53]"},{"why":"Restormer is the runner-up Attenuation Net backbone in the RACDI backbone comparison.","marker":"[48]"}],"fun_headline_variants":["Re-degrade the restored image to judge its fidelity","Forget the original: compare restored to degraded image","Blind restoration? Use the degraded photo as the ground truth","New IQA for blind restoration: consistency with degraded image","Fidelity of blind restoration: re-degrade and compare"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that DISDCD's subjective choices—'which re-degraded restoration looks closer to the degraded image'—are a valid operationalization of fidelity; if human fidelity judgment includes anything beyond low-frequency closeness to the degraded image, the experimental support weakens.","fun_headline_variants_meta":{"raw":{"variants":["Re-degrade the restored image to judge its fidelity","Forget the original: compare restored to degraded image","Blind restoration? Use the degraded photo as the ground truth","New IQA for blind restoration: consistency with degraded image","Fidelity of blind restoration: re-degrade and compare"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001088,"raw_usage":{"total_tokens":4578,"prompt_tokens":1010,"completion_tokens":3568,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":626,"completion_tokens_details":{"reasoning_tokens":3488}},"tokens_in":626,"tokens_out":3568,"duration_ms":22888,"temperature":1.0,"reasoning_tokens":3488,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:15:06.999844+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive test would be a preference study where raters choose which restoration is more faithful to the original scene, with the pristine original shown as reference, among restorations that have identical low-frequency consistency with the degraded image but different high-frequency realism; if raters consistently prefer one, CDI misses part of fidelity. A simpler check is to compute CDI on a dataset whose ground truth was collected against the pristine original rather than by re-degrading restorations.","supporting_citations":[{"cited_title":"Sheikh and A.C","cited_arxiv_id":null,"evidence_quote":"Supplies the wavelet-domain two-step degradation model and HVS mutual-information formulas used for noise removal and noise-equivalent attenuation."},{"cited_title":"Real-esrgan: Training real-world blind super-resolution with pure synthetic data","cited_arxiv_id":null,"evidence_quote":"Provides the Real-ESRGAN second-order degradation pipeline used to generate combined degradation and DISDCD test images."},{"cited_title":"Ntire 2017 challenge on single image super-resolution: Dataset and study","cited_arxiv_id":null,"evidence_quote":"DIV2K is the source image pool for DISDCD and for the blur-sensitivity demonstration that motivates CDI."},{"cited_title":"Efros, Eli Shecht- man, and Oliver Wang","cited_arxiv_id":null,"evidence_quote":"LPIPS is a perceptual baseline and the 2AFC protocol for scoring DISDCD and computing human accuracy."},{"cited_title":"Bovik, H.R","cited_arxiv_id":null,"evidence_quote":"SSIM is the main full-reference baseline whose scores increase under added blur, motivating the shift to CDI."},{"cited_title":"Simoncelli","cited_arxiv_id":null,"evidence_quote":"DISTS is the second perceptual full-reference baseline compared on DISDCD."},{"cited_title":"Practical blind denoising via swin-conv-unet and data synthesis, 2022","cited_arxiv_id":null,"evidence_quote":"Supplies the SCUNet backbone trained as the Attenuation Net for the reference-agnostic RACDI."},{"cited_title":"Restormer: Efficient transformer for high-resolution image restoration","cited_arxiv_id":null,"evidence_quote":"Restormer is the runner-up Attenuation Net backbone in the RACDI backbone comparison."}],"review_version":1}