{"id":"eb2f6012-ad99-4ae1-b5cf-5339a1c95414","arxiv_id":"2504.12222","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A two-stage network that fuses video codec motion vectors and residuals with a diffusion model improves no-reference perceptual scores on GoPro and DVD, while PSNR and SSIM fall far below existing methods.","lead":"CPGD-Net is a video deblurring method that extracts motion vectors and coding residuals from a video codec and uses them, together with a pretrained diffusion model, to generate sharper frames. It reports large gains on no-reference perceptual metrics, but the tradeoff is a substantial drop in traditional pixel-level metrics.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The perceptual-quality claim is supported only by no-reference metrics and omits the sole prior diffusion-based video deblurring baseline, leaving SOTA status unverified.","rationale":"The reader's weakest assumption about coding residuals being reliable blur indicators is a legitimate robustness concern for the proposed mechanism, but I see the empirical evaluation as more load-bearing for the central claim. Even if the coding-residual assumption holds perfectly, the paper's evidence for state-of-the-art perceptual quality would still be incomplete because it relies on no-reference metrics, omits the only closely related diffusion-based baseline, and contains an apparent reporting anomaly in Table I. The reader's rationale already mentions the no-reference metric issue and the missing VD-Diff comparison, so there is partial agreement, but the reader's formal weakest_assumption field points to the coding-residual mechanism rather than the evaluation validity. My recommended disposition remains conditional: the core idea is plausible and worth publishing only after the evaluation is strengthened with human perceptual judgments, a VD-Diff comparison, and correction or explanation of the repeated PSNR value. Therefore the reader's CONDITIONAL verdict is unchanged, though my stated reason differs from the reader's stated weakest assumption.","tokens_in":12137,"tokens_out":5581,"duration_ms":63338,"concrete_test":"Run a pre-registered two-alternative forced-choice human preference study on 100 matched frames from GoPro and DVD, comparing Ours against the best non-diffusion baseline (BSST or RVRT) and against VD-Diff, while also recomputing PSNR/SSIM/LPIPS and verifying the Table I entries. If Ours is not preferred at a statistically significant level, or if the NIQE/MUSIQ advantage reverses when recomputed, the central claim of state-of-the-art perceptual quality is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that CPGD-Net achieves state-of-the-art perceptual quality, with NIQE and MUSIQ improvements of roughly 27-30% and 7-28% on GoPro and DVD. The evidence in Table I is dominated by no-reference metrics: NIQE and MUSIQ. The method also shows much lower PSNR and SSIM than every baseline (e.g., 26.80 dB versus 35.98 dB on GoPro), which the authors attribute to the perception-distortion tradeoff. However, no-reference metrics are not a substitute for human perceptual evaluation in generative restoration; NIQE assumes natural scene statistics and can be improved by adding high-contrast but hallucinated textures, and MUSIQ is trained on human ratings of natural images, not deblurring outputs. Without a human study, the reported NIQE/MUSIQ gains do not establish that the outputs are actually preferred by viewers or are more faithful to the deblurred scene. The paper also cites VD-Diff [22], the only prior diffusion-based video deblurring method, but does not compare against it in Table I, so the claim of state-of-the-art performance is not supported against the most relevant baseline. Finally, Table I reports identical PSNR values (26.80) on two different datasets, with very close SSIM values (0.7808 and 0.7809); this is likely a reporting error and further weakens confidence in the quantitative results. Code and data are promised but not available, preventing independent reproduction. These issues are load-bearing because the central claim is an empirical claim about perceptual superiority.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CPGD-Net, a two-stage video deblurring framework that combines coding priors from video compression (motion vectors and coding residuals) with a pretrained Stable Diffusion model via a ControlNet-like module. Stage 1 (CPFP) aligns features using motion vectors and residual-derived attention masks, producing an initial restoration. Stage 2 (CPC) injects coding priors into the diffusion model's attention to guide high-detail synthesis. The authors claim state-of-the-art perceptual quality, reporting NIQE reductions of 27–30% and MUSIQ gains of 7–28% on GoPro and DVD, while accepting substantially lower PSNR/SSIM than all baselines. Ablations show the contribution of each component and compare motion vectors with RAFT optical flow.","tokens_in":12464,"tokens_out":3141,"duration_ms":33032,"significance":"The core idea of exploiting codec-side motion vectors and coding residuals as priors for video deblurring is novel and potentially practical, since MVs come at no extra computational cost. The two-stage design, ablations, and the decision to open-source code and data are positive features. If the perceptual claim were backed by proper human evaluation and the missing diffusion-based baseline were included, this would be a useful contribution to the video deblurring literature. However, the current evaluation is not sufficient to support the central claim.","major_comments":[{"comment":"The reported PSNR values for GoPro and DVD are identical (26.80) and the SSIM values are nearly identical (0.7808 and 0.7809) across two different datasets, which is statistically implausible. This suggests a reporting error, perhaps from copying a single result or applying the same test set. Because the paper's quantitative narrative depends heavily on these numbers, the authors must correct Table I and rerun all metrics, or provide an explanation for the coincidence.","section":"Table I"},{"comment":"The central claim of 'state-of-the-art perceptual quality' rests entirely on the no-reference metrics NIQE and MUSIQ, with no user study. Given that PSNR drops by roughly 9 dB relative to the best baselines (e.g., 26.80 vs 35.98 on GoPro) and SSIM drops by about 0.2, the burden of proof is high. No-reference metrics are known to favor generative outputs that add high-contrast but possibly hallucinated texture, and neither NIQE nor MUSIQ is calibrated for video deblurring outputs. The authors should provide a human perceptual study (e.g., paired comparison or MOS) and also report full-reference perceptual metrics such as LPIPS on the restored content to demonstrate that the improvements are genuine and not merely artifacts of no-reference scoring.","section":"Section IV.C, Table I"},{"comment":"The paper cites VD-Diff [22] as the only prior diffusion-based video deblurring method, but does not compare against it in Table I or elsewhere. If the paper claims state-of-the-art performance, omitting the most relevant baseline leaves the claim unverified. The authors should add VD-Diff (and any other diffusion-based video deblurring baselines) to the quantitative and qualitative comparisons, and discuss any differences in the evaluation protocol that would justify exclusion.","section":"Section II.A, Table I"},{"comment":"The assumption that coding residuals identify blur, occlusion, and texture variation is load-bearing for the CPFA masks (Eqs. 2–3) and the CPC attention mask (Eq. 6), yet it is never tested across different codecs, bitrates, or encoder configurations. All experiments use VVC with an unspecified configuration, and the residual distribution is strongly shaped by quantization, rate control, and GOP structure. As presented, the claimed benefit of coding priors may not generalize. The authors should at least specify the VVC configuration used, and ideally ablate over a range of bitrates and codecs (e.g., H.264/H.265) to show that the attention masks remain beneficial and do not amplify coding artifacts.","section":"Section III.A, Section IV.B"}],"minor_comments":[{"comment":"The text prompt in the figure contains a typo: 'High quility' should be 'High quality'.","section":"Figure 2"},{"comment":"The caption reads 'Detile of the CPFA block' and should be 'Detail of the CPFA block'.","section":"Figure 4 caption"},{"comment":"The phrase 'pre-train diffusion models' should be 'pre-trained diffusion models'.","section":"Section II.B"},{"comment":"The abstract reports 'up to 30% improvement in IQA metrics' while the conclusion gives specific 28%/7% MUSIQ and 30%/27% NIQE gains; these numbers should be made consistent in one place.","section":"Abstract and Section V.G"},{"comment":"The abbreviation 'W/O DONATES WITHOUT' is unclear; presumably it means 'w/o denotes without', but it reads awkwardly and should be rewritten.","section":"Tables II and III"},{"comment":"The implementation refers to a 'spaced DDPM sampling schedule' with citation [42], but it is not clear whether the final inference uses DDIM, DDPM, or another sampler. Please specify the exact sampling algorithm and number of steps used for the reported results.","section":"Section IV.B"},{"comment":"Several equations and sentences lack spaces, e.g., 'V andR' and 'F L i'; please fix these formatting issues for readability.","section":"Section III.C"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a novel and interesting combination of codec priors with diffusion-based generation, and the method has potential. However, the evaluation is the main obstacle: the perceptual claim is supported only by no-reference metrics, the sole diffusion-based video deblurring baseline is omitted, and Table I contains a suspicious identical-PSNR value that must be resolved. These are fixable with additional experiments and a human study, so I am not recommending rejection, but the revision needs to be substantial."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the core idea—feeding VVC motion vectors and coding residuals into a two-stage deblurring network, then conditioning a frozen Stable Diffusion with a ControlNet that takes those priors—is genuinely new and worth someone's time. The paper's engineering is coherent; the ablations show CPFP and CPC do something. But the evidence for the headline claim is weak: no-reference metrics only, no comparison against VD-Diff, and a PSNR table that looks wrong.\n\nWhat's good: using coding residuals as attention masks is a reasonable way to point a deblurring network at motion-blurred regions, and the design of CPFP follows directly from that. The comparison against optical flow (Table III) is a fair and useful way to test the coding-prior assumption, and the result—MVs nearly match RAFT-20 at zero FLOPs—is the most credible result in the paper. The CPC module on top of stage-one output is a standard ControlNet-style architecture, but the masking via MVs/CRs is a legitimate twist. No signs of data invention; the method is described well enough to reproduce once code ships.\n\nThe soft spots are the ones your stress-tester named, and they land. The SOTA claim rests on NIQE and MUSIQ; no user study, no comparison to VD-Diff, and the paper's own Figure 6 shows the perceptual-vs-distortion gap, which doesn't absolve them from testing perception directly. The identical PSNR 26.80 on GoPro and DVD is a red flag—almost certainly a reporting error, and combined with the very low PSNR/SSIM it undermines confidence in the table. The 'coding residual indicates blur' assumption is plausible but untested across codecs/bitrates; one encoding configuration from VVC is thin. Also the extra dataset curated with CodeFormer-degraded clips, and the reliance on it to avoid overfitting, is underexplained—could easily leak or bias the perceptual metrics.\n\nWho is this for: researchers working on compressed-domain video restoration or generative restoration. A serious referee should see it, because the combination is new and the engineering is honest. But I'd want the authors to add a human study, include VD-Diff, fix the table, and test at least two bitrates before publication.","headline":"A genuinely novel combination of codec side-information and diffusion conditioning, undercut by an evaluation that cannot support the SOTA claim.","tokens_in":13000,"tokens_out":1550,"would_cite":false,"duration_ms":16604,"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":"CPGD-Net claims that pairing video codec motion vectors and residuals with a pretrained diffusion model yields state-of-the-art perceptual quality in video deblurring.","keywords":["video deblurring","motion vectors","coding residuals","diffusion model","ControlNet","perceptual quality","no-reference image quality assessment","video coding priors"],"falsifier":"Encode the same test videos with different codecs (for example HEVC versus VVC) and widely different bitrates, then compare the residual maps with known blur masks; if residual maps change with compression decisions far more than with blur content, the attention masks would amplify coding artifacts instead of restoring detail. A concrete check would compute the overlap between large-residual regions and ground-truth blur regions on the GoPro dataset under VVC at QP 22 versus QP 37.","tokens_in":11950,"feed_emoji":"🎞️","tokens_out":3452,"duration_ms":32802,"temperature":0.7,"pith_summary":"The paper proposes a two-stage video deblurring pipeline, CPGD-Net, that feeds motion vectors and coding residuals extracted from a compressed video stream into a network along with the decoded frames. The first stage uses these codec priors for efficient inter-frame alignment and to build attention masks; the second stage feeds the restored frames plus the same priors into a pretrained latent diffusion model adapted through a control network. The authors claim this produces state-of-the-art perceptual quality, with 27-30 percent better NIQE and 7-28 percent better MUSIQ scores than previous methods on GoPro and DVD. They argue that pixel-level metrics like PSNR and SSIM are lower because generative detail synthesis is penalized by full-reference metrics, and they advocate no-reference quality assessment instead. The paper's core insight is that video compression already computes exactly the motion and residual information a deblurring network needs, and a generative prior can then fill in realistic texture.","feed_headline":"Codec priors plus diffusion lift video deblurring quality scores","feed_subtitle":"Method reports 27-30% better NIQE and 7-28% better MUSIQ on GoPro/DVD, at the cost of lower PSNR.","key_machinery":"The load-bearing objects are the two codec priors: motion vectors $V_{t-1\\to t}$, which warp previous-frame features, and coding residuals $R_{t-1\\to t}$, normalized to $[0,1]$, which are added as skip connections to deformable-convolution offsets and masks in the Coding-Prior Feature Alignment (CPFA) block. In the Coding-Prior Controlled Generation (CPC) module, a control network takes $V$ and $R$, converts them through a linear layer into an attention mask $A$, and modulates the query of self-attention in the denoising U-Net, steering generation toward motion-affected regions. The pretrained diffusion model supplies realistic texture priors; the codec priors supply motion and attention.","core_discovery":"On its own terms, the central discovery is that coding residuals - the parts of a compressed frame that motion compensation could not predict - act as a reliable map of where blur, occlusion, and texture change live, and that this map can be converted into attention masks that guide both a deformable alignment stage and a diffusion-based refinement stage. The paper shows that replacing optical flow with free motion vectors is competitive while costing zero added computation, and that removing residual masks hurts quality. It also demonstrates that the diffusion stage only works well when the first alignment stage is present and when additional high-quality training data is used; without these, the generator overfits and reproduces reference frames with insufficient detail.","pith_inferences":["The residual-as-blur-map assumption likely depends on the encoder, bitrate, and GOP structure; testing on HEVC or AVC or at low bitrates would show whether compression decisions, not blur, shape the masks.","An ablation that replaces coding-residual masks with random or edge-based masks would isolate whether the coding prior contributes beyond generic attention.","The same two-stage recipe - codec priors for alignment and a control-conditioned diffusion for detail - may transfer to compressed-domain video super-resolution or video frame interpolation.","A user study rather than no-reference metrics would test whether the claimed perceptual gains survive human judgment, since NIQE and MUSIQ are only proxies."],"forward_implications":["Video deblurring can be performed on compressed streams with no extra optical-flow computation, since motion vectors come from the decoder.","Coding residuals provide a cheap prior for where deblurring should focus, potentially useful for other restoration tasks.","Perceptual quality metrics NIQE and MUSIQ can diverge sharply from PSNR and SSIM; evaluation of generative deblurring should consider both.","Diffusion-based refinement requires a strong initial restoration stage, otherwise blurred regions remain unresolved.","Additional high-quality training data beyond the deblurring benchmarks is needed to prevent generative overfitting."],"supporting_citations":[{"why":"Provides the pretrained latent diffusion model that supplies generative priors for realistic detail synthesis.","marker":"[19]"},{"why":"ControlNet is the conditional control mechanism adapted into the CPControlNet design.","marker":"[20]"},{"why":"DiffBIR's IRControlNet serves as the baseline and initialization for the coding-prior control network.","marker":"[28]"},{"why":"RAFT optical flow is the baseline that motion vectors are compared against in the ablation studies.","marker":"[53]"},{"why":"The VVC codec used to extract motion vectors and coding residuals from the training and test videos.","marker":"[6]"},{"why":"BasicVSR++ is the propagation and alignment architecture that the CPFA blocks draw on.","marker":"[9]"},{"why":"RVRT is a transformer-based baseline whose performance CPGD-Net compares against.","marker":"[11]"},{"why":"BSSTNet is a blur-aware transformer baseline that the perceptual comparisons include.","marker":"[13]"}],"fun_headline_variants":["Coding residuals guide diffusion for sharper video deblurring","Free motion vectors plus diffusion improve deblurring quality","Video deblurring scores jump with codec-informed diffusion","Deblurring networks tap codec priors and generative diffusion","Residual masks from codecs steer diffusion for realistic deblurring"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole pipeline depends on the assumption that regions with large coding residuals reliably mark where motion blur and texture loss occur, even though coding residuals are also shaped by encoder decisions like quantization and rate control.","fun_headline_variants_meta":{"raw":{"variants":["Coding residuals guide diffusion for sharper video deblurring","Free motion vectors plus diffusion improve deblurring quality","Video deblurring scores jump with codec-informed diffusion","Deblurring networks tap codec priors and generative diffusion","Residual masks from codecs steer diffusion for realistic deblurring"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00031,"raw_usage":{"total_tokens":1733,"prompt_tokens":874,"completion_tokens":859,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":490,"completion_tokens_details":{"reasoning_tokens":771}},"tokens_in":490,"tokens_out":859,"duration_ms":8734,"temperature":1.0,"reasoning_tokens":771,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:34:38.239600+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Encode the same test videos with different codecs (for example HEVC versus VVC) and widely different bitrates, then compare the residual maps with known blur masks; if residual maps change with compression decisions far more than with blur content, the attention masks would amplify coding artifacts instead of restoring detail. A concrete check would compute the overlap between large-residual regions and ground-truth blur regions on the GoPro dataset under VVC at QP 22 versus QP 37.","supporting_citations":[{"cited_title":"Recurrent video restoration transformer with guided deformable attention,","cited_arxiv_id":null,"evidence_quote":"RVRT is a transformer-based baseline whose performance CPGD-Net compares against."},{"cited_title":"Blur-aware spatio-temporal sparse transformer for video deblurring,","cited_arxiv_id":null,"evidence_quote":"BSSTNet is a blur-aware transformer baseline that the perceptual comparisons include."}],"review_version":1}