{"id":"35846cb4-742f-4cc5-8711-b9a066e2e0c9","arxiv_id":"1908.07404","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Blind deconvolution is solved by alternating gradient descent in the latent spaces of pretrained image and blur-kernel generators, with a slack variant that relaxes the image constraint.","lead":"This paper deblurs images without knowing the blur by running gradient descent on the latent codes of two pretrained networks, one that generates sharp images and one that generates blur kernels. It reports higher PSNR than classic deblurring baselines on face, shoe, and digit datasets, and shows the method degrades gracefully as noise grows without retraining.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The strongest Table 1 numbers come from DDS, not the plain generative-prior objective Eq. (4); DDS adds a free pixel image and a TV term, and no ablation isolates whether the generative image prior drives the gain.","rationale":"The reader's weakest assumption identifies the generator-range limitation, which is real and explicitly acknowledged. The concern raised here is different but related: the paper's most impressive quantitative results are produced by the hybrid objective Eq. (5), which includes a free image variable and a hand-crafted TV prior. Without an ablation that removes the generative image term, the paper does not establish that pretrained generative models, rather than the added TV regularization and extra degrees of freedom, are responsible for beating DeblurGAN and classical baselines on CelebrA and Shoes. On SVHN the pure method Eq. (4) does win, so the generative-prior idea has some support, but the broad central claim as stated in the abstract and introduction requires the stronger datasets to be attributed correctly. This concern is empirically addressable and does not by itself disprove the method, so it strengthens the case for a conditional verdict rather than moving it to accept or reject. The reader's conditional verdict is therefore unchanged.","tokens_in":9641,"tokens_out":8751,"duration_ms":93765,"concrete_test":"Run the Eq. (5) pipeline on the CelebA and Shoes test sets with the generative image constraint ablated (τ=0), keeping G_K, TV, and all hyperparameters otherwise fixed; then separately run it with ρ=0. If the τ=0 variant stays within roughly 0.5 dB PSNR of the reported DDS values, the generative image prior is not the driver of the reported gains; if the ρ=0 variant collapses while the τ=0 variant does not, the TV term is the key component.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table 1 shows that the headline claim 'minimizing Eq. (4) yields results that beat classical priors and compete with or exceed DeblurGAN' is not supported by the pure method on the harder datasets: Deep Deblur (Eq. 4) is below DeblurGAN on Shoes (21.20 vs 21.84 dB) and CelebA (21.11 vs 24.01 dB). The values that exceed DeblurGAN are produced by Deep Deblur with Slack, whose objective (5) introduces an unconstrained image variable i, a total-variation prior ρ||i||_tv, and three extra weights (τ, ζ, ρ). No ablation is reported for Eq. (5): τ=0 removes the generative image constraint, ρ=0 removes TV, and ζ=0 removes the in-range consistency term. Consequently, the paper has not demonstrated that the pretrained generative model is the cause of the improvement over classical priors; a TV-regularized blind deconvolution using the same kernel generator and random restarts might perform comparably on these 80-image benchmarks. Since the paper's novelty claim is specifically about generative priors, this missing attribution is load-bearing.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a blind image deconvolution method that regularizes the ill-posed inverse problem by constraining the recovered image and blur kernel to the ranges of two pretrained generative models. The authors formulate a latent-space objective (Eq. 4) and minimize it by alternating gradient descent with random restarts, calling this Deep Deblur. To handle datasets where the generator range is insufficient, they introduce Deep Deblur with Slack (DDS, Eq. 5), which adds a free pixel-domain image, a range-error penalty, and a total-variation prior. Experiments on SVHN, Shoes, and CelebA report PSNR/SSIM values against classical deblurring priors and end-to-end networks, plus noise and large-blur robustness experiments. The paper's central claims are that this is the first use of pretrained generative models for blind image deblurring and that the proposed approach outperforms existing methods without retraining for different noise levels.","tokens_in":9924,"tokens_out":5410,"duration_ms":50332,"significance":"If the central claims are fully supported, the paper proposes a conceptually interesting way to use pretrained generative models for blind deconvolution, combining latent-space optimization with a slack variable to mitigate generator range limitations. The idea is timely and the results on the tested datasets are promising, especially the DDS variant. However, the paper's experimental evidence currently has a load-bearing attribution gap: the best numbers come from the slack objective that mixes generative and classical priors, and the contribution of the generative image prior is not isolated. The paper also provides no quantitative evaluation of kernel estimation accuracy, even though joint image-kernel recovery is a stated contribution, and it defers all hyperparameters and algorithmic details to a missing supplementary. These issues need to be addressed before the claims can be accepted at face value.","major_comments":[{"comment":"The quantitative claim that the proposed approach outperforms all competitors on Shoes and CelebA is carried by DDS (Eq. 5), not by the pure generative-prior objective (Eq. 4). On Shoes, Deep Deblur attains 21.20 dB versus 21.84 dB for DeblurGAN and 24.76 dB for CNN; on CelebA it attains 21.11 dB versus 24.01 dB for DeblurGAN. DDS adds a free pixel-domain image variable i, a total-variation term ρ||i||_tv, and three weights τ, ζ, ρ beyond Eq. (4). No ablation is reported (e.g., τ=0, ρ=0, or a TV-only blind deconvolution using the same kernel generator). Without such an ablation, the observed gains cannot be attributed to the pretrained image generator, which is the paper's central novelty. This is a load-bearing omission.","section":"Section 3.2.2, Table 1"},{"comment":"The paper's stated contribution is recovering 'true image and blur kernel' (Section 1), and Eq. (4) and (5) optimize both latent variables for the image and the kernel. However, all quantitative results in Table 1 and Figure 9 are image PSNR/SSIM only. No metric for blur kernel estimation accuracy (e.g., kernel MSE, normalized cross-correlation, or visual comparison of recovered kernels) is reported. The claim of joint image-kernel recovery is therefore not substantiated by the experiments.","section":"Section 2 and Section 3"},{"comment":"The choice of all free parameters (γ, λ, τ, ζ, ρ), the gradient step sizes, and the random-restart schedule are deferred to a supplementary material that is not included with the manuscript. Because the method is a non-convex optimization with tuned hyperparameters, omitting these values makes the experiments non-reproducible and prevents the reader from assessing the sensitivity of the results to parameter choices. The supplementary material should be provided, or at minimum a table of all hyperparameter values and a description of the random-restart procedure should be added.","section":"Section 3.1"}],"minor_comments":[{"comment":"The manuscript contains several typographical and grammatical errors, for example 'with in' in the contributions list, 'T echnology' in the author affiliation, and 'these comparison' in Section 3.2.2. A careful proofreading pass is needed.","section":"Throughout"},{"comment":"The random-restart criterion is described as restarting when the measurement loss 'does not reduce sufficiently after reasonably many iterations'; this is too vague to reproduce. Please specify the exact threshold and iteration counts.","section":"Section 2"},{"comment":"The noise and blur robustness curves are presented as averages, but no error bars or confidence intervals are shown. Given that Table 1 reports only 80 test images per dataset, adding variance information would help assess the statistical reliability of the comparisons.","section":"Section 3.2.3, Figure 9"},{"comment":"The CNN baseline from [13] is designed for text deblurring, but it is evaluated on Shoes and CelebA. Please clarify whether this network was retrained or used off-the-shelf, and whether any domain adaptation was performed.","section":"Section 3.1"},{"comment":"The SVHN row has no DDS entry, and the text explains the choice, but this should also be stated in the table caption for clarity.","section":"Section 3.2.2, Table 1"}],"recommendation":"major_revision","confidential_remarks":"The main issue for the editor is the attribution gap: the paper's headline results are produced by a mixed generative-plus-classical objective, and without an ablation (e.g., setting τ=0 or ρ=0) the paper does not demonstrate that the generative image prior is responsible for the gains. This is fixable with additional experiments, as are the missing kernel metrics and the absent supplementary details. I would be willing to review a revised version that addresses these points."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper you asked about is a mixed bag: the core idea is a straightforward extension of the generative-prior inverse-problem framework to blind deconvolution, and it is honestly written, but the best quantitative results come from the \"slack\" variant, not from the generative prior itself, and no ablation isolates whether the generator is what does the work. I agree with the stress-test note: it holds up.\n\nWhat is new: two pretrained generators, one for images and one for blur kernels, optimized in latent space with alternating gradient descent and random restarts. That is a natural adaptation of Bora et al. and Hand et al., and the paper says so. The range-image analysis in Section 3.2.1 is genuinely useful: it gives a way to separate generator-range error from deblurring error, and the authors are explicit that Deep Deblur tends to recover the nearest range image rather than the true image on CelebA. The noise-robustness and large-blur experiments are a good addition, and the PGGAN experiment at 128x128 suggests the method scales if the generator is expressive.\n\nThe soft spots are real. Table 1 shows Deep Deblur (Eq. 4) is below DeblurGAN on Shoes (21.20 vs 21.84 dB) and CelebA (21.11 vs 24.01 dB). The numbers that beat the end-to-end baselines come from Deep Deblur with Slack (Eq. 5), which introduces a free pixel image, a TV regularizer, and three extra weights but no ablation. It is possible that a TV-regularized blind deconvolution with the same kernel generator and restarts would perform comparably, which would undercut the paper's central claim. The authors need to run the tau=0, rho=0, zeta=0 variants and show the generator is load-bearing. Also, the promised supplementary material is missing, no code is released, the quantitative results are on 80 images per dataset with no error bars, and hyperparameters are per-dataset. These are fixable, but they are not minor when reproducibility is the standard.\n\nBottom line: the paper is a credible extension with an honest discussion of its own limitation, and I would send it to a serious referee, but I would not accept it in current form. The missing ablation and missing artifacts are exactly what a strong referee should request. For my own work, I would cite the range-image analysis as a useful evaluation tool. It is a good reading-group paper for discussing how to attribute gains in generative-prior methods.","headline":"A credible but incomplete extension of generative priors to blind deconvolution; the best numbers come from the slack variant and the paper never isolates the generative prior's contribution.","tokens_in":10428,"tokens_out":3153,"would_cite":true,"duration_ms":30803,"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 blind image deblurring can be solved by alternating gradient descent in the latent spaces of two pretrained generative models—one for sharp images and one for blur kernels—so that no end-to-end retraining is required.","keywords":["blind image deblurring","blind deconvolution","generative priors","latent space optimization","alternating gradient descent","GAN","VAE","image restoration"],"falsifier":"Take a test image $i_{\\text{test}}$ far from the generator range, compute its nearest range image $i_{\\text{range}}$, blur both with the same kernel and noise to form $y_{\\text{test}}$ and $y_{\\text{range}}$, then run Deep Deblur on each. If the output from $y_{\\text{test}}$ is no closer to $i_{\\text{test}}$ than $i_{\\text{range}}$ already is, the recovery claim fails for out-of-range images.","tokens_in":9434,"feed_emoji":"🖼️","tokens_out":10859,"duration_ms":91188,"temperature":0.7,"pith_summary":"Blind image deblurring is a severely ill-posed inverse problem, and this paper argues that the missing regularization can come from pretrained generative models rather than hand-crafted image priors or end-to-end training. The proposal is to search the latent spaces of two fixed generators—one trained on sharp images, the other on blur kernels—by alternating gradient descent on the objective $\\|y - G_I(z_i)\\otimes G_K(z_k)\\|^2 + \\gamma\\|z_i\\|^2 + \\lambda\\|z_k\\|^2$. The paper claims this recovers both the true image and the blur kernel to within the range of the respective generators, and that a slack variant that allows the image to leave the generator range, penalized by total variation, fixes the main failure mode on complex datasets. If true, this offers a way to deblur under large blur and heavy noise without retraining the network for new noise levels.","feed_headline":"Latent-space search deblurs images without retraining","feed_subtitle":"Two pretrained generators—for sharp images and blur kernels—turn blind deconvolution into alternating gradient descent.","key_machinery":"The load-bearing objects are the two pretrained generators $G_I: \\mathbb{R}^l \\to \\mathbb{R}^n$ and $G_K: \\mathbb{R}^m \\to \\mathbb{R}^n$, whose ranges act as the feasible sets for the sharp image and the blur kernel. The algorithm is alternating gradient descent on the latent codes $(z_i, z_k)$ to minimize the data-fidelity term $\\|y - G_I(z_i)\\otimes G_K(z_k)\\|^2$ plus $\\ell^2$ penalties on the codes, with random restarts to escape poor local minima. The named algorithms are Deep Deblur (strictly constrained to the generator ranges) and Deep Deblur with Slack (DDS), which adds an intermediate image $i$ tied to $G_I(z_i)$ by $\\|i - G_I(z_i)\\|^2$ and regularized by total variation, allowing recovery outside the generator range. The key diagnostic quantity is the range error $\\|i_{\\text{test}} - i_{\\text{range}}\\|$, the distance from a true test image to the nearest image the generator can produce.","core_discovery":"The central claim is that the range of a pretrained image generator is a strong enough prior to disambiguate blind deconvolution. Given a blurry observation $y$, the paper solves $(\\hat z_i,\\hat z_k) = \\arg\\min_{z_i,z_k}\\|y - G_I(z_i)\\otimes G_K(z_k)\\|^2 + \\gamma\\|z_i\\|^2 + \\lambda\\|z_k\\|^2$ with random restarts, where $G_I$ and $G_K$ are fixed pretrained generators for images and blur kernels. The authors report that this 'Deep Deblur' procedure yields sharp deblurred images that beat classical priors such as dark-channel and extreme-channel methods, and that on rich datasets the 'Deep Deblur with Slack' variant—which adds a total-variation term and lets the image deviate from the generator range—outperforms end-to-end networks such as DeblurGAN. The paper frames the approach as the first use of pretrained generative models for blind image deblurring.","pith_inferences":["This suggests the same alternating latent-space scheme could apply to other blind inverse problems, such as blind super-resolution or blind inpainting, whenever one generator for clean signals and one for the degradation are available.","The range-error bottleneck points to a concrete upgrade path: replacing the GAN/VAE generator with a more expressive generative prior, such as a diffusion model, should improve deblurring on complex natural image classes.","A testable practical corollary is that the slack parameter $\\tau$ in the DDS objective should scale with the measured range error of the image generator class; the paper does not provide such a rule."],"forward_implications":["Blind deblurring becomes a fixed-cost optimization over pretrained models: the same image and blur generators handle different blur sizes and noise levels without retraining.","The quality of deblurring is bounded by the expressiveness of the image generator; images inside the generator range are recovered much more accurately than arbitrary images.","On complex image classes where the generator range is incomplete, the slack variant that allows out-of-range images should be used instead of strict range constraint.","End-to-end deblurring networks lose their advantage as blur size or noise grows, while the generative-prior approach degrades more gracefully."],"supporting_citations":[{"why":"Establishes the generative-prior inverse-problem approach that this method adapts to blind deblurring.","marker":"[2]"},{"why":"Supplies the motion-blur generation strategy used to synthesize the blur-kernel dataset.","marker":"[3]"},{"why":"Provides the end-to-end CNN deblurring baseline compared in the experiments.","marker":"[13]"},{"why":"Provides the progressive GAN used to demonstrate deblurring at higher resolution.","marker":"[15]"},{"why":"Provides the end-to-end adversarial baseline (DeblurGAN) compared in the experiments.","marker":"[18]"},{"why":"Supplies the dark-channel classical prior baseline.","marker":"[25]"},{"why":"Supplies the DCGAN architecture used for the image generators on Shoes and CelebA.","marker":"[28]"},{"why":"Defines SSIM, the quantitative similarity metric used for evaluation.","marker":"[33]"}],"fun_headline_variants":["Pretrained GANs deblur blind images via latent search","Blind deblurring by alternating latent-space descent","GAN priors fix blind deconvolution without retraining","Two generators, one deblur: latent optimization wins","Blind deconvolution solved by pretrained generative priors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes the true sharp image lies close to the range of the pretrained image generator; if the generator cannot represent the test image, the deblurred output collapses toward the nearest generator-range image rather than the true image.","fun_headline_variants_meta":{"raw":{"variants":["Pretrained GANs deblur blind images via latent search","Blind deblurring by alternating latent-space descent","GAN priors fix blind deconvolution without retraining","Two generators, one deblur: latent optimization wins","Blind deconvolution solved by pretrained generative priors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000363,"raw_usage":{"total_tokens":1922,"prompt_tokens":878,"completion_tokens":1044,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":494,"completion_tokens_details":{"reasoning_tokens":960}},"tokens_in":494,"tokens_out":1044,"duration_ms":7847,"temperature":1.0,"reasoning_tokens":960,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:18:30.848633+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a test image $i_{\\text{test}}$ far from the generator range, compute its nearest range image $i_{\\text{range}}$, blur both with the same kernel and noise to form $y_{\\text{test}}$ and $y_{\\text{range}}$, then run Deep Deblur on each. If the output from $y_{\\text{test}}$ is no closer to $i_{\\text{test}}$ than $i_{\\text{range}}$ already is, the recovery claim fails for out-of-range images.","supporting_citations":[{"cited_title":"Modeling the performance of image restora- tion from motion blur","cited_arxiv_id":null,"evidence_quote":"Supplies the motion-blur generation strategy used to synthesize the blur-kernel dataset."},{"cited_title":"Convolutional neural networks for direct text deblurring","cited_arxiv_id":null,"evidence_quote":"Provides the end-to-end CNN deblurring baseline compared in the experiments."},{"cited_title":"DeblurGAN: Blind Motion Deblurring Using Conditional Adversarial Networks","cited_arxiv_id":"1711.07064","evidence_quote":"Provides the end-to-end adversarial baseline (DeblurGAN) compared in the experiments."},{"cited_title":"Blind image de- blurring using dark channel prior","cited_arxiv_id":null,"evidence_quote":"Supplies the dark-channel classical prior baseline."}],"review_version":1}