{"id":"720d1e64-6472-404a-8963-2636d99c41c8","arxiv_id":"2507.09285","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A generative blur-kernel prior with a learned initializer improves blind motion deblurring across DIP, VDIP, and diffusion-based methods, including non-uniform cases.","lead":"The authors train a GAN to learn the shapes of motion blur kernels and a second network that predicts a starting kernel from a blurry image. This plug-and-play module improves several existing deblurring methods and reports state-of-the-art scores on standard benchmarks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper never isolates the proposed generative manifold constraint from the learned kernel initializer, so the central claim that the manifold alleviates initialization sensitivity is not directly tested.","rationale":"The reader's weakest assumption (transfer from synthetic uniform pretraining to real/non-uniform kernels) is plausible, but the paper explicitly pretrains on kernels of the Lai type and thanks the provider of the Lai-kernel generation code, so the Lai benchmark is at least partially in-distribution. The more fundamental gap is that the mechanism itself is never isolated. The experiments vary initialization and kernel parameterization simultaneously, so the reported gains are consistent with the learned initializer alone. This does not refute the method, but it means the central assertion that the generative latent manifold is the active ingredient is not yet established. The missing ablated control is a concrete, low-cost experiment, and the absent supplementary limitations statement prevents readers from checking the authors' own caveats. The verdict should therefore remain CONDITIONAL: the empirical claim may hold, but the mechanistic claim requires the ablations described above before acceptance.","tokens_in":25796,"tokens_out":8723,"duration_ms":107092,"concrete_test":"Run a three-way ablation on the Lai synthetic benchmark and the non-uniform dataset: (a) baseline DIP/VDIP with random kernel init; (b) same baseline, but initialize the kernel as G_k(E(y)) and then optimize the kernel freely with no generator constraint; (c) full GLKM. Add (d) full GLKM but replace E(y) with a random latent code. If (b) ≈ (c), the generator constraint is inert; if (d) ≈ (c), the initializer is unnecessary; if both matter, report the incremental gains. Also report mean ± std over 5 random seeds for each variant to quantify sensitivity to initialization.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central assertion is that combining a GAN-based kernel generator (manifold constraint) with a learned kernel initializer alleviates sensitivity to kernel initialization. The experiments only compare each baseline with its default (usually random) kernel initialization against GLKM, which changes two things at once: the initialization is now the ResNet encoder's prediction, and the kernel is optimized in the generator's feature space instead of freely. No experiment holds one factor fixed. In particular, there is no baseline 'DIP + same E(y) initialization but free kernel optimization' and no 'DIP-GLKM with random latent initialization instead of E(y)', so the contribution of the manifold constraint is unknown. The paper's own Fig. 1 demonstrates baseline sensitivity only anecdotally; no variance-over-initializations comparison is reported. Without these ablations, the reported gains could be due entirely to the supervised initializer, and the claim that the compact latent kernel manifold is what helps is unsupported. This is aggravated by the conclusion's pointer to 'limitations discussed in the supplementary material,' which is not included in the arXiv text, so the authors' own caveats are unavailable for assessment.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GLKM, a plug-and-play generative latent kernel prior for blind motion deblurring (BMD). The method pre-trains a GAN-based kernel generator on synthetic motion blur kernels and a kernel initializer (a ResNet encoder) that maps a blurry image to the latent code of the generator. During deblurring, the kernel is optimized in the first-layer feature space of the generator rather than in the low-dimensional latent code, and is initialized by the encoder's prediction. GLKM is integrated into four existing BMD frameworks—DIP, VDIP, BIRD, and BlindDPS—and is extended to non-uniform BMD via the SVOLA formulation with a global kernel as a coherent initialization. Experiments on synthetic datasets, the Lai benchmark, AFHQ-dog, CelebAHQ, and ImageNet report consistent improvements over the corresponding baselines and state-of-the-art results on Lai's non-uniform dataset.","tokens_in":26045,"tokens_out":4151,"duration_ms":49947,"significance":"If the reported gains hold, the plug-and-play nature of GLKM is a practical strength: it improves several deep-prior BMD baselines without retraining their image generators, and it extends to non-uniform blur without adding hand-crafted kernel priors. The integration with multiple image priors (DIP, VDIP, diffusion-based BIRD and BlindDPS) and the public code release are commendable. However, the central conceptual claim—that constraining the kernel to a compact latent manifold alleviates the sensitivity to kernel initialization—is not directly evidenced by the experiments as presented. The paper's main contribution is therefore promising but under-validated; the missing ablations and variance analysis are needed to support the headline assertion.","major_comments":[{"comment":"The experiments never isolate the proposed generative manifold constraint from the learned kernel initializer. In every comparison, the baseline uses its default (usually random) kernel initialization and free-form kernel optimization, while GLKM uses both the encoder's predicted initialization and optimization in the generator's first-layer feature space w_k. To support the claim that the compact latent manifold itself stabilizes BMD, the paper should include at least two ablations: (i) baseline DIP/VDIP with the same E(y) initialization but optimizing the kernel directly in the original kernel space, and (ii) GLKM with a random latent initialization instead of E(y). Without these, the reported gains could be attributed entirely to the supervised initializer, and the central assertion is unsupported. In addition, the paper asserts in Sec. V-A that optimizing w_k is better than optimizing z_k, but no experiment compares these two optimization spaces.","section":"Sec. V-A, Eqs. (5)-(6), Tables I-II"},{"comment":"The paper's motivation is that BMD is extremely sensitive to the initial blur kernel, yet no quantitative evidence of reduced sensitivity is provided. Fig. 1 shows two anecdotal runs of SelfDeblur with different random initializations, but the paper reports no variance or repeated-run statistics for any baseline or for GLKM. Since the proposed method is specifically designed to stabilize the solution, the authors should report mean and standard deviation (or at least per-run results) over multiple random restarts for both the baselines and GLKM on a subset of the test images. A single run per method is insufficient to substantiate a sensitivity-reduction claim.","section":"Sec. III, Fig. 1"},{"comment":"The kernel generator and initializer are pre-trained only on synthetic uniform blur kernels (following [20] or [30]) and then applied directly to the non-uniform Lai dataset, whose kernels are constructed from inertial sensor recordings of real camera motion. The paper asserts transferability without presenting evidence that the learned generator's support includes such kernels. If the synthetic manifold does not contain the true test-time kernels, the manifold constraint would bias the solution toward incorrect kernels. The authors should provide a kernel-space transfer analysis, for example by projecting the ground-truth Lai kernels through the pre-trained generator and measuring reconstruction error, or by comparing estimated kernels with ground truth on the non-uniform dataset.","section":"Sec. VII, Pre-training Setup; Sec. IV-B"},{"comment":"The manuscript repeatedly defers essential content to a supplementary file that is not present in the arXiv submission: the detailed VDIP-GLKM implementation and BlindDPS integration are described only in the supplementary material, and the conclusion states that limitations are discussed in the supplementary material. As a result, key parts of the method and the authors' own caveats are unavailable for assessment. The supplementary should be made available, or the relevant details and limitations should be summarized in the main text.","section":"Sec. VIII; Sec. V-A.1"},{"comment":"There is a conspicuous anomaly in the non-uniform results: VDIP-Extreme-GLKM achieves SSIM 0.370 on the synthetic non-uniform dataset (Table V), far below DIP-GLKM (0.807) and even below VDIP-Sparse-GLKM (0.584). In Table VI, VDIP-Extreme-GLKM is consistently worse than VDIP-Std-GLKM across all categories, which contradicts the uniform experiment where VDIP-Extreme-GLKM is competitive with or better than other variants. This pattern suggests a possible bug or reporting error in the non-uniform VDIP-Extreme implementation; please verify and correct, or explain why this variant degrades in the non-uniform setting.","section":"Tables V and VI"}],"minor_comments":[{"comment":"The hyperparameter λ is fixed to 0.1, but the paper provides no sensitivity analysis for this choice. A brief experiment varying λ would help establish robustness.","section":"Eq. (4)"},{"comment":"The notation in Algorithm 1 is confusing: line 10 sets θ0 = θt−1_E, but the outer loop uses θt−1_E as the current parameter. Please clarify the iteration indices and the role of the inner L-loop.","section":"Algorithm 1"},{"comment":"The caption contains a typo: 'Esitimated Kernels' should be 'Estimated Kernels'.","section":"Fig. 3 caption"},{"comment":"The comparison list for non-uniform deblurring includes methods up to 2023; since the paper claims state-of-the-art performance, it would be useful to compare against more recent non-uniform deblurring methods from 2024–2025 if any are available.","section":"Sec. VII, Comparison Methods"},{"comment":"The paper states that baselines were run with 'officially default settings' or tuned for best performance, but it does not specify which setting was used for each comparison method. This information is needed for reproducibility.","section":"Sec. VI-A, Implementation Details"}],"recommendation":"major_revision","confidential_remarks":"The paper is an extension of the authors' ECCV 2024 work [37], and the main text references that work as the source of the DIP-based framework. The novelty relative to [37] appears to be the extension to diffusion-based image priors (BIRD, BlindDPS) and to non-uniform BMD; this should be stated more clearly so that the incremental contribution can be judged. A more serious editorial issue is that the supplementary material—referenced for VDIP/BlindDPS details, additional visual comparisons, and the paper's own limitations—is not included in the submitted arXiv text. For a journal submission, the supplementary should be available to the reviewers. The central missing ablation (isolating the initializer from the manifold constraint) is fixable and should be required before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a credible extension, not a new paradigm. The core pieces—GAN kernel generator plus latent encoder initializer—come from the authors' ECCV 2024 paper. What's new is showing they can be dropped into diffusion-based methods (BIRD, BlindDPS) and into non-uniform SVOLA deblurring, with consistent gains. That is useful and largely convincing. The code is public, and the evaluation is broad: uniform synthetic, Lai synthetic and real, non-uniform synthetic and Lai, plus three 256x256 diffusion benchmarks.\n\nWhat it does well: the plug-and-play story holds up. DIP-GLKM, BIRD-GLKM, and BlindDPS-GLKM all beat their baselines, and the non-uniform results are striking because the unmodified DIP/VDIP baselines collapse. I believe the combination of a learned initialization and a kernel-manifold constraint helps.\n\nSoft spots. The central mechanistic claim is untested. The paper says the compact latent kernel manifold alleviates sensitivity to kernel initialization, but every comparison changes two things at once: the start point (random -> E(y)) and the optimization space (free kernel -> first-layer feature of G_k). There is no ablation holding one fixed—no DIP with E(y) initialization but free kernel optimization, and no GLKM with a random latent code instead of E(y). Fig. 1 is two runs, which is anecdotal. Given that the paper's own motivation is variance across initializations, the lack of error bars or repeated-run numbers is a real gap. The stress-test note is right; this is the main soft spot.\n\nSecond, the transfer concern is fair but secondary. The generator and initializer are trained only on synthetic uniform kernels (Sec. IV-B and VII pre-training), yet applied to Lai's inertial-sensor non-uniform kernels. The real-data and non-uniform results suggest it transfers in practice, but the uniform synthetic test kernels come from the same generation recipe as the training kernels, so the uniform numbers probably overstate generalization.\n\nThird, the paper defers to the supplementary for implementation details and says 'limitations discussed in the supplementary material' in the conclusion, but that material is not in the arXiv text. That is not a scientific flaw, but it makes the authors' own caveats unavailable.\n\nWho it's for: people working on deep-prior blind deblurring and plug-and-play priors for inverse problems. I'd send it to review. A serious referee should ask for the missing ablations, error bars, and the supplementary; none of that requires rejecting the work.","headline":"A credible plug-and-play extension of the authors' ECCV kernel-prior work, with broad experiments, but the central claim that the manifold constraint rather than the learned initializer drives the gains is not actually tested.","tokens_in":26566,"tokens_out":3597,"would_cite":false,"duration_ms":44340,"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":"The paper claims that a pretrained GAN-based kernel prior, combined with a learned kernel initializer, confines blind motion deblurring to a compact latent kernel manifold, stabilizes kernel estimation, and improves several existing…","keywords":["blind motion deblurring","generative kernel prior","kernel initializer","latent kernel manifold","deep generative prior","non-uniform deblurring","GAN-based kernel generator","plug-and-play prior"],"falsifier":"Run the pretrained kernel generator and initializer on a test set of camera-shake kernels that were recorded by inertial sensors and compare recovered kernels with the recorded trajectories; if initialization outside the synthetic manifold produces kernels that stay on the manifold but miss the true trajectory, and deblurring quality degrades accordingly, the transfer assumption fails. A simpler check is to ablate the encoder by restarting optimization from random latent codes drawn from the generator's prior and measure whether the reported gains disappear.","tokens_in":25593,"feed_emoji":"🌀","tokens_out":7078,"duration_ms":79286,"temperature":0.7,"pith_summary":"The paper tries to show that the main obstacle in deep-prior blind motion deblurring is not the image prior but the way the blur kernel is initialized and searched. It proposes a two-part pretrained module: a GAN-based generator that encodes motion-blur kernels in a compact latent space, and an encoder that predicts a good latent code directly from the blurry image. Plugging this module into existing DIP-based and diffusion-based deblurring methods is said to improve their accuracy and stability, and to make non-uniform deblurring work without extra handcrafted kernel priors. A sympathetic reader would care because the module is presented as a reusable kernel prior rather than a full new deblurring pipeline, so its claimed benefit would compound across many existing methods.","feed_headline":"Latent kernel prior lifts four blind motion deblurring baselines","feed_subtitle":"A pretrained GAN kernel generator and encoder start the search near true blur, cutting random-start failures.","key_machinery":"The load-bearing object is the latent kernel manifold: the image of a pre-trained GAN-based kernel generator. Because the generator compresses kernels into a low-dimensional latent space, optimizing the latent code (or, better, the first-layer feature map of the generator) keeps the search on plausible kernel shapes. The kernel initializer, an encoder mapping the blurry image to a latent code, gives the optimizer a warm start in that manifold. Together they turn kernel estimation from an unconstrained high-dimensional problem into a constrained one.","core_discovery":"The central claim is that the instability of deep-prior blind motion deblurring comes from searching for the blur kernel in the full, high-dimensional kernel space, where random starting points fall into bad local minima. To fix this, the paper pre-trains a GAN that maps a low-dimensional latent code to realistic motion-blur kernels and an encoder that maps the blurry image to a good latent code for that GAN. During deblurring the kernel is optimized only inside the generator's output space, initialized by the encoder, which confines the solution to a compact latent kernel manifold. The paper reports that this constraint improves four existing deblurring frameworks built on deep image priors or diffusion models, and that the same pretrained kernel prior transfers to non-uniform blur where many spatially varying kernels must be estimated together.","pith_inferences":["If the latent-kernel-prior recipe is as general as it looks, it should transfer to other blind inverse problems with structured unknown operators, such as unknown super-resolution kernels or atmospheric point-spread functions, as long as a generator can be trained on the operator family; the paper demonstrates only motion blur.","The first-layer feature-map optimization suggests a tunable trade-off between manifold restriction and search flexibility; one could test how results degrade as the optimized layer moves deeper or the latent dimension shrinks, which would map where the sweet spot lies.","Because the generator is frozen during optimization, any test-time kernel outside the learned manifold cannot be recovered; an extension would be to detect low-confidence encoder initializations and adaptively widen the search or fine-tune the generator."],"forward_implications":["Existing deep-prior deblurring methods can be upgraded by swapping in the pretrained kernel prior, gaining several dB of PSNR and better perceptual scores on synthetic and real blurred images.","Random-restart sensitivity that produced widely divergent kernels in DIP-based methods is largely removed, making outputs more reproducible across runs.","The same pretrained components extend to non-uniform blur, where many spatially varying kernels are estimated jointly, without adding handcrafted kernel regularizers.","Large kernels near 75 by 75 pixels become tractable because optimization is confined to a compact manifold instead of the full kernel space.","Diffusion-based deblurring also improves when the kernel comes from the latent prior rather than from a separately trained network or handcrafted regularization."],"supporting_citations":[{"why":"Supplies the DIP-based blind deblurring formulation whose kernel-initialization sensitivity is the paper's motivating problem.","marker":"[25]"},{"why":"Provides the variational DIP baseline and its variants that the plug-and-play kernel prior is tested on.","marker":"[26]"},{"why":"The diffusion-based baseline that uses a diffusion kernel prior, which GLKM is compared against and also integrated with.","marker":"[30]"},{"why":"The diffusion-inversion deblurring baseline into which the proposed latent kernel modeling is embedded.","marker":"[31]"},{"why":"The earlier work that introduced the generative kernel prior and latent encoder, which this paper extends to more image priors and non-uniform blur.","marker":"[37]"},{"why":"Supplies the benchmark and the kernel-generation trajectories used for training and evaluation, including real camera shakes.","marker":"[38]"},{"why":"Provides the physical motion-kernel synthesis mechanism used to build the pretraining kernel collection.","marker":"[20]"},{"why":"The style-based GAN architecture that motivates optimizing the kernel generator's first-layer feature map instead of its latent input.","marker":"[79]"}],"fun_headline_variants":["Generative latent kernel model fixes blind deblurring initialization","Pretrained kernel generator steers blind deblurring to better starts","GAN latent space narrows blind deblurring kernel search","Plug-and-play latent kernel prior improves blind deblurring","Kernel manifold from GAN cuts blind deblurring start sensitivity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pretrained kernel manifold, learned from synthetic uniform blur kernels, is assumed to contain the real, spatially varying test-time kernels, so no retraining or adaptation is needed when the module is transferred.","fun_headline_variants_meta":{"raw":{"variants":["Generative latent kernel model fixes blind deblurring initialization","Pretrained kernel generator steers blind deblurring to better starts","GAN latent space narrows blind deblurring kernel search","Plug-and-play latent kernel prior improves blind deblurring","Kernel manifold from GAN cuts blind deblurring start sensitivity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000891,"raw_usage":{"total_tokens":3837,"prompt_tokens":934,"completion_tokens":2903,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":2813}},"tokens_in":550,"tokens_out":2903,"duration_ms":25278,"temperature":1.0,"reasoning_tokens":2813,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:59:29.788991+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the pretrained kernel generator and initializer on a test set of camera-shake kernels that were recorded by inertial sensors and compare recovered kernels with the recorded trajectories; if initialization outside the synthetic manifold produces kernels that stay on the manifold but miss the true trajectory, and deblurring quality degrades accordingly, the transfer assumption fails. A simpler check is to ablate the encoder by restarting optimization from random latent codes drawn from the generator's prior and measure whether the reported gains disappear.","supporting_citations":[{"cited_title":"Neural blind deconvolution using deep priors,","cited_arxiv_id":null,"evidence_quote":"Supplies the DIP-based blind deblurring formulation whose kernel-initialization sensitivity is the paper's motivating problem."},{"cited_title":"Blind image deconvolution using variational deep image prior,","cited_arxiv_id":null,"evidence_quote":"Provides the variational DIP baseline and its variants that the plug-and-play kernel prior is tested on."},{"cited_title":"Parallel diffusion models of operator and image for blind inverse problems,","cited_arxiv_id":null,"evidence_quote":"The diffusion-based baseline that uses a diffusion kernel prior, which GLKM is compared against and also integrated with."},{"cited_title":"Blind image restoration via fast diffusion inversion,","cited_arxiv_id":null,"evidence_quote":"The diffusion-inversion deblurring baseline into which the proposed latent kernel modeling is embedded."},{"cited_title":"Blind image deconvolution by generative-based kernel prior and initializer via latent encoding,","cited_arxiv_id":null,"evidence_quote":"The earlier work that introduced the generative kernel prior and latent encoder, which this paper extends to more image priors and non-uniform blur."},{"cited_title":"A comparative study for single image blind deblurring,","cited_arxiv_id":null,"evidence_quote":"Supplies the benchmark and the kernel-generation trajectories used for training and evaluation, including real camera shakes."},{"cited_title":"A style-based generator architecture for generative adversarial networks,","cited_arxiv_id":null,"evidence_quote":"The style-based GAN architecture that motivates optimizing the kernel generator's first-layer feature map instead of its latent input."}],"review_version":1}