{"id":"0594d343-a782-4516-963b-4e29e9777581","arxiv_id":"1908.02197","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A zero-shot neural optimization, using a Deep Image Prior network for the sharp image and a fully connected Softmax network for the blur kernel, achieves competitive state-of-the-art blind deconvolution on benchmarks.","lead":"This paper introduces a neural network method that removes blur from a single photo without needing any training examples of blurry and sharp image pairs. It models the sharp image and the blur kernel as outputs of two small networks and fits them to the observed blurry image directly.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The objective in Eq. (6) admits a trivial delta-kernel solution whose loss is no larger than the true solution's, so the reported gains depend on early stopping at T=5,000 and implicit optimization bias rather than on a demonstrated deep prior.","rationale":"The reader's conditional verdict already identifies the architectural-prior assumption and lack of convergence guarantee as the weakest point. My stress-test sharpens this: the regularized objective itself has a trivial delta-kernel global solution with loss no larger than the true solution, so the method's success is logically tied to an unconverged, initialization-dependent trajectory. This is a real flaw in the paper's framing as a deep-prior optimization, but it does not by itself overturn the reported empirical benchmark gains, which are reproducible from the released code. The appropriate response is to keep the conditional verdict and add an explicit condition: demonstrate that results are stable under longer optimization and multiple seeds, or disclose early stopping as the actual mechanism. Hence no change to the reader's verdict is needed.","tokens_in":14052,"tokens_out":7637,"duration_ms":85779,"concrete_test":"Run the released SelfDeblur code on the Levin benchmark (or one representative image) with T extended to 50,000 and with 5 random seeds for the zx perturbation, recording per-iteration fidelity, TV, kernel error ratio, and PSNR. If the loss continues to decrease while kernel error ratio worsens or PSNR degrades after an intermediate peak, the T=5,000 results are an early-stopping artifact; if the solution remains near the true kernel for all seeds, the concern is refuted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The critical weakness is not just the absence of a convergence proof; Eq. (6) is demonstrably biased away from the true solution. For the trivial solution k = δ, x = y, the fidelity term is zero and the loss is λ·TV(y). For the true solution, the loss is λ·TV(x) (ignoring noise). Since y = k⊗x with a nonnegative unit-mass kernel cannot have larger total variation than x, TV(y) ≤ TV(x), so the delta solution has loss no larger than the true solution. The FCN Gk can approximate δ arbitrarily closely (Softmax with one dominant logit), and the overparameterized DIP Gx can fit y, so the global infimum of Eq. (6) is ≤ λ·TV(y). Thus minimizing Eq. (6) does not favor the true (x,k). The paper's assertion in Sec. 3.3 that 'the powerful modeling capacity of Gk and Gx is beneficial to avoid trivial delta kernel solution' is unsupported: the benchmark numbers at T=5,000 come from an unconverged trajectory, and no mechanism other than early stopping and initialization separates SelfDeblur from the delta solution. That makes the claimed generalization to new kernels and noise levels precarious.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SelfDeblur, a zero-shot blind deconvolution method that replaces handcrafted MAP priors with two untrained generative networks: a DIP-style asymmetric autoencoder Gx for the latent clean image and a fully-connected network Gk with a Softmax output layer for the blur kernel. The method optimizes the unconstrained objective in Eq. (6) by ADAM, with joint optimization as the default, and reports state-of-the-art average PSNR/SSIM/error-ratio results on the Levin and Lai benchmarks as well as favorable qualitative results on real blurry images. The authors argue that the generative-network priors and joint optimization avoid the trivial delta-kernel solution, making non-blind deconvolution optional.","tokens_in":14400,"tokens_out":5238,"duration_ms":55490,"significance":"If the empirical results hold, the paper makes a useful conceptual contribution: it extends the deep-image-prior idea to blind deconvolution and shows that untrained generators can serve as image and kernel priors without any training data. The paper's strengths include public code, a clear architecture ablation, and quantitative comparisons on standard benchmarks. However, the theoretical explanation for avoiding the trivial delta-kernel solution is not supported by the objective in Eq. (6), and the empirical robustness claims are weakened by the absence of seed sensitivity and hyperparameter sensitivity analyses. The contribution is therefore significant but conditional on additional analysis and on a corrected or carefully qualified claim about the optimization mechanism.","major_comments":[{"comment":"The objective in Eq. (6) is mathematically biased toward the trivial delta-kernel solution, so the claim that 'the powerful modeling capacity of Gk and Gx is beneficial to avoid trivial delta kernel solution' is unsupported. For any nonnegative unit-mass kernel k, convolution is a convex combination of shifted copies of x, and hence TV(k*x) <= TV(x). For the trivial solution k=delta, x=y, the loss is lambda*TV(y). For the true solution, the loss is lambda*TV(x) + ||k*x - y||^2 >= lambda*TV(x) >= lambda*TV(y). Since the Softmax FCN can approximate a delta kernel arbitrarily closely and the overparameterized DIP can fit y, the global infimum of Eq. (6) is no larger than the value at the true solution. The reported results at T=5,000 therefore depend on early stopping and on the optimization trajectory, not on a demonstrated property of the generative priors. The authors should either add a mechanism that actually excludes the delta solution, provide an analysis of the optimization bias that explains why early stopping selects the true solution, or empirically isolate the role of initialization and stopping time across many seeds.","section":"Sec. 3.3, Eq. (6)"},{"comment":"All ablations are conducted on a single dataset (Levin et al.) and no seed or initialization sensitivity is reported, even though the method randomly samples zx and zk with a fixed seed and also perturbs zx randomly at each iteration. The central empirical claim of robustness would be considerably strengthened by reporting means and standard deviations over at least several random seeds on both the Levin and Lai benchmarks, including error ratios and failure cases. Without this, the average PSNR of 33.07 on Levin could reflect a favorable initialization or an early-stopping artifact.","section":"Sec. 4.1, Tables 1-2"},{"comment":"The TV regularization weight is hand-chosen as lambda = 0.1*sigma, with sigma estimated by [54], and no sensitivity analysis is provided. Since the TV term is explicitly introduced to handle noise and to prevent overfitting, the benchmark results depend on this choice. The paper should report how PSNR, SSIM, and kernel error vary with lambda on the Levin and Lai datasets, and justify the linear schedule rather than treating it as a fixed default.","section":"Sec. 3.2 and Sec. 4.2"}],"minor_comments":[{"comment":"Line 8 of Algorithm 2 states 'x = G^T_x(z_k)' but should presumably read 'k = G^T_k(z_k)'.","section":"Algorithm 2"},{"comment":"The caption says the generative networks capture 'the deep priors of blur kernel and latent clean image', but Gx models the image and Gk models the kernel; the order appears reversed.","section":"Figure 2 caption"},{"comment":"The text says zx is sampled with fixed random seed 0 but then states that zx is perturbed randomly at each iteration following [45]; the relationship between the fixed seed and the per-iteration perturbation should be clarified.","section":"Sec. 4.1"},{"comment":"There are several typographical errors, including 'equalitly', 'emperically', 'Quantitavie', 'Machaeli', and 'Perroe'; these should be corrected in a revision.","section":"General"},{"comment":"The claim that SelfDeblur 'makes the first attempt' at state-of-the-art blind deconvolution with deep networks is stronger than the evidence in the paper supports and should be softened.","section":"Sec. 1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be based on a previously published conference version, so for a journal submission the bar for analysis is higher than for the original conference paper. The main risk is that the Sec. 3.3 claim about avoiding the delta solution is directly contradicted by the form of Eq. (6), and a careful reviewer could use this to undermine the contribution. The empirical results are promising, so I would not reject, but the revision must address the optimization-bias issue head-on and provide the missing robustness experiments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth your time if you work on blind deconvolution or deep image priors. It extends DIP to blind deconvolution by pairing a DIP-style autoencoder for the clean image with a small FCN + Softmax network for the kernel, and it shows that joint optimization of both networks on a single blurry image can beat several MAP baselines on the Levin and Lai benchmarks. The code is available, the experiments are clearly described, and the benchmark numbers are genuinely competitive. That part is real.\n\nThe soft spot is not just missing theory; the objective in Eq. (6) is demonstrably biased toward the trivial solution. If you set the kernel to a delta and the image to the blurry input, the fidelity term is zero and the loss is lambda times TV(y). For the true kernel and sharp image, the loss is lambda times TV(x). Since blur decreases total variation, TV(y) <= TV(x), the delta solution has loss no larger than the true solution. The FCN with Softmax can approximate a delta, and the DIP can overfit y, so the global infimum of Eq. (6) is at most lambda times TV(y). The claim in Sec. 3.3 that the powerful modeling capacity avoids the trivial delta solution is unsupported. The good results at T=5,000 come from an unconverged trajectory and implicit optimization bias, not from a prior that prefers the true kernel. That is a load-bearing caveat.\n\nThere are also smaller issues: no error bars or seed sensitivity analysis, the TV weight is hand-set as 0.1 times the estimated noise level, and all ablations are on one dataset. The paper also does not discuss related deep-generative blind deconvolution work that appeared around the same time; that is worth a citation check.\n\nThat said, the paper is honest about being zero-shot self-supervised, and the empirical contribution is substantial enough that the missing theoretical grounding does not sink it. For a reader who wants to see a concrete DIP-based blind deconvolution recipe, this is a solid reference. For a referee, the delta-solution issue needs to be confronted directly, and the authors should report variance across seeds and iterations. I would send it to peer review rather than desk reject; with revisions addressing the objective gap, it could be a good paper.","headline":"Useful empirical paper on DIP-based blind deconvolution, but the objective provably favors the delta kernel; the reported gains rest on early stopping, not a demonstrated deep prior.","tokens_in":14842,"tokens_out":1413,"would_cite":true,"duration_ms":16713,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that blind deconvolution can be reduced to fitting two randomly initialized neural networks to a single blurry image, with the network architectures themselves acting as priors that outperform handcrafted MAP-based…","keywords":["blind deconvolution","zero-shot learning","deep image prior","generative networks","blur kernel estimation","self-supervised learning","image restoration","neural optimization"],"falsifier":"Take a synthetic blurry image with a known kernel, run the method from many different random seeds, and compare the estimated kernels to the ground truth; the architectural-prior claim predicts consistently accurate kernels, while the weakest-assumption failure would show delta-kernel or high-variance collapses on some seeds. A simpler check is whether the reported error ratio on the small benchmark degrades sharply when the image generator alone is replaced by a plain autoencoder without skip connections.","tokens_in":13884,"feed_emoji":"🖼️","tokens_out":7879,"duration_ms":80127,"temperature":0.7,"pith_summary":"The paper tries to show that blind deconvolution—recovering a sharp image and the blur kernel from one blurry photo—can be solved without any training data and without handcrafted image priors, by letting the architecture of two small neural networks supply the prior. The networks are randomly initialized, then optimized on the single blurry image so that the convolution of the generated sharp image and generated kernel reproduces the observation. On two standard blind-deconvolution benchmarks the method reports average PSNR 33.07 with error ratio 1.1968 on the first and average PSNR 20.97 on the second, ahead of the MAP-based methods it compares against. If correct, this means a generic untrained network structure can carry enough information to regularize a severely ill-posed inverse problem.","feed_headline":"Two untrained networks beat tuned deblurring methods on benchmarks","feed_subtitle":"A randomly initialized autoencoder and a small network learn the sharp image and blur kernel from the blurry photo alone.","key_machinery":"The central object is the pair of generative networks Gx and Gk inserted into the forward model y = k ⊗ x + n. Instead of regularizing x and k directly, the objective regularizes the network parameters through the architectures: Gx is a deep image prior, an untrained skip-connected encoder–decoder that preferentially produces natural-image-like structures, while Gk is a deliberately simple fully-connected network whose output layer uses Softmax, making the non-negativity and sum-to-one kernel constraints automatic. The fixed random input vectors zx and zk turn blind deconvolution into an unconstrained optimization over network parameters only, solved by joint ADAM updates with an added TV term whose weight is set by the estimated noise level.","core_discovery":"The paper's central claim is that both unknowns in blind deconvolution—the latent clean image and the blur kernel—can be represented as outputs of generative networks with fixed random inputs, and that fitting those networks to the blurry observation, plus a small total-variation term, estimates both unknowns without paired training data. The image network is an asymmetric autoencoder with skip connections and a sigmoid output; the kernel network is a fully-connected network whose final Softmax layer guarantees the kernel is non-negative and sums to one. The paper argues that this unconstrained neural optimization avoids the trivial delta-kernel solution that can trap conventional MAP methods, and that jointly updating both networks with ADAM performs better than alternating their updates. The reported results include average PSNR 33.07 and error ratio 1.1968 on the smaller benchmark, and average PSNR 20.97 on the larger benchmark, which the paper presents as improvements over the compared state of the art.","pith_inferences":["Beyond the paper: this two-network factorization is a template for other blind inverse problems with a convolutional forward model, such as joint demosaicking with unknown blur or super-resolution with an unknown downsampling kernel; the same architecture-as-prior logic should be testable there without changing the method's structure.","Beyond the paper: since the paper fixes one random seed and one iteration budget, the relative roles of architecture prior versus early stopping are unresolved; a seed-averaging or early-stopping experiment would separate these and could make the method's behavior more predictable.","Beyond the paper: the comparison against learned deep deblurring is limited to one retrained network on one benchmark; a broader comparison on larger, non-uniform, or saturated kernels would clarify whether the zero-shot prior generalizes beyond the tested distributions."],"forward_implications":["A single blurry image suffices: no paired training data or learned deblurring network is needed, so the method can in principle be applied to cameras or image types for which no training set exists.","Because the kernel network's Softmax output enforces non-negativity and sum-to-one at every iteration, the projection and delayed-normalization steps that MAP methods rely on are unnecessary.","The latent image produced by Gx already has visually favorable texture, making the second non-blind deconvolution stage optional rather than compulsory.","The TV weight is tied to the estimated noise level through $\\lambda = 0.1\\sigma$, so the method explicitly adapts its prior strength when the blurry image is also noisy."],"supporting_citations":[{"why":"Supplies the deep image prior architecture that Gx is built on; the paper's core premise is that this untrained network captures natural-image statistics.","marker":"[45]"},{"why":"The coupled deep-image-prior baseline that the paper tests and finds inadequate for modeling blur kernels, motivating the separate FCN for Gk.","marker":"[6]"},{"why":"Provides the benchmark dataset and the analysis showing MAP methods can collapse to the trivial delta kernel, the failure mode SelfDeblur claims to avoid.","marker":"[19]"},{"why":"Provides the non-blind deconvolution method used after kernel estimation and the error ratio metric used for evaluation.","marker":"[20]"},{"why":"A main state-of-the-art MAP baseline that SelfDeblur compares against on benchmarks and real-world images.","marker":"[29]"},{"why":"Establishes that optimization details such as delayed kernel normalization determine MAP success, providing the background against which the unconstrained joint optimization is proposed.","marker":"[32]"},{"why":"The ADAM optimizer used in both alternating and joint optimization algorithms to update the network parameters.","marker":"[13]"}],"fun_headline_variants":["Zero-shot deblurring: two untrained nets beat tuned methods","SelfDeblur: fitting random networks to blurry photos","No paired data needed: neural optimization deblurs images","Untrained autoencoder and FCN jointly deblur benchmarks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a randomly initialized, untrained network architecture supplies a prior strong enough that minimizing the reconstruction loss on one blurry image drives the solution to the true kernel and sharp image rather than to a trivial delta kernel or to noise fitting.","fun_headline_variants_meta":{"raw":{"variants":["Zero-shot deblurring: two untrained nets beat tuned methods","SelfDeblur: fitting random networks to blurry photos","No paired data needed: neural optimization deblurs images","Untrained autoencoder and FCN jointly deblur benchmarks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000229,"raw_usage":{"total_tokens":1514,"prompt_tokens":1013,"completion_tokens":501,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":629,"completion_tokens_details":{"reasoning_tokens":429}},"tokens_in":629,"tokens_out":501,"duration_ms":6059,"temperature":1.0,"reasoning_tokens":429,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:51:53.340854+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a synthetic blurry image with a known kernel, run the method from many different random seeds, and compare the estimated kernels to the ground truth; the architectural-prior claim predicts consistently accurate kernels, while the weakest-assumption failure would show delta-kernel or high-variance collapses on some seeds. A simpler check is whether the reported error ratio on the small benchmark degrades sharply when the image generator alone is replaced by a plain autoencoder without skip connections.","supporting_citations":[{"cited_title":"Ulyanov, A","cited_arxiv_id":null,"evidence_quote":"Supplies the deep image prior architecture that Gx is built on; the paper's core premise is that this untrained network captures natural-image statistics."},{"cited_title":"Gandelsman, A","cited_arxiv_id":null,"evidence_quote":"The coupled deep-image-prior baseline that the paper tests and finds inadequate for modeling blur kernels, motivating the separate FCN for Gk."},{"cited_title":"Levin, Y","cited_arxiv_id":null,"evidence_quote":"Provides the benchmark dataset and the analysis showing MAP methods can collapse to the trivial delta kernel, the failure mode SelfDeblur claims to avoid."},{"cited_title":"Levin, Y","cited_arxiv_id":null,"evidence_quote":"Provides the non-blind deconvolution method used after kernel estimation and the error ratio metric used for evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"A main state-of-the-art MAP baseline that SelfDeblur compares against on benchmarks and real-world images."},{"cited_title":"Perrone and P","cited_arxiv_id":null,"evidence_quote":"Establishes that optimization details such as delayed kernel normalization determine MAP success, providing the background against which the unconstrained joint optimization is proposed."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The ADAM optimizer used in both alternating and joint optimization algorithms to update the network parameters."}],"review_version":1}