{"id":"46823574-d386-4007-a7db-c58ffb855c85","arxiv_id":"2412.09193","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A short-exposure guided diffusion network detects blurry regions, restores them with a masked guided filter, and uses a diffusion prior to generate realistic, sharp moving objects.","lead":"A smartphone deblurring method that pairs the usual blurry photo with a second, short-exposure shot of the same scene, using the sharp details from the second shot to guide a diffusion model in cleaning up the blurry moving objects. The result is a system that removes local motion blur while keeping the background sharp, and it reports better perceptual-quality scores than existing deblurring networks on public and newly captured data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The real-world gains on LMBD may reflect a domain shift between synthetic short-exposure training frames and real smartphone captures; without alignment analysis or a real-pair comparison, the role of short-exposure guidance is unverified.","rationale":"The reader's weakest assumption identifies the same load-bearing point: the synthetic short-exposure images used in training must faithfully represent real smartphone short-exposure captures, including alignment and sharp structure of the moving objects. This assumption is central because the guided restoration in Eqs. (4)-(6) and the ExpBFusion module explicitly inject R as structural guidance; if the real R is misaligned or has a different noise process, the guidance can corrupt the restored image. The paper provides no quantitative bridge between synthetic and real short-exposure distributions, and the LMBD dataset details are relegated to an unavailable supplemental description. This is not an internal inconsistency, but it means the real-world component of the central claim is unverified. My concern does not move the reader's verdict: the paper should remain CONDITIONAL, with the condition being a demonstration that synthetic-to-real transfer holds, ideally by comparing against a model trained with real short-exposure pairs and by reporting alignment statistics.","tokens_in":10921,"tokens_out":11092,"duration_ms":115298,"concrete_test":"Retrain ExpRDiff from scratch using only the Section 4.1 synthetic short-exposure generation on ReLoBlur, then evaluate on the LMBD test set; fine-tune the same architecture on the real LMBD training pairs and re-evaluate. If the synthetic-only model's NIQE/CLIPIQA+/MUSIQ-PAQ2PIQ on LMBD drop substantially relative to the real-pair model, the synthetic short-exposure assumption is load-bearing. As a minimal additional check, report the mean optical-flow alignment error between the long- and short-exposure frames in LMBD static regions; if this exceeds a few pixels, Eqs. (4)-(6) are operating on misaligned references and alignment must be modeled.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1 generates short-exposure training images by converting sharp frames to HSV, reducing brightness, and adding random noise. At inference, however, the short-exposure image R is a real smartphone capture from a different sensor pipeline, and Eqs. (4)-(6) use R as the structural reference for H. If the real R is not aligned with the long-exposure frame (parallax, slight motion, or exposure differences) or has a different noise/color distribution, the learned guided filter and ExpBFusion can copy wrong structure into the blurry regions. The paper does not report any alignment preprocessing for LMBD, does not quantify the distribution shift between synthetic and real R, and does not publicly describe the capture protocol or the 10,380-image training set beyond a pointer to the supplemental material. Table 2's large NIQE gains could therefore be produced by the SDXL prior hallucinating plausible texture rather than by the short-exposure guidance, so the claim that short-exposure images facilitate accurate reconstruction is not yet supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ExpRDiff, a three-module pipeline for local motion deblurring: a context-based local blur detection module that classifies non-overlapping image patches with Gumbel-Softmax; a blur-aware guided image restoration step that applies a masked guided filter using the detected confidence map and a short-exposure image R; and a short-exposure guided diffusion model built on SDXL, in which multi-scale features from R and the blur map are fused into the diffusion U-Net features via an attention module (ExpBFusion). The method is trained on ReLoBlur with synthetic short-exposure images generated by darkening sharp frames in HSV space and adding noise, and on a newly captured real dataset, LMBD. Quantitative results are reported on ReLoBlur and LMBD using only the no-reference metrics NIQE, CLIPIQA+, and MUSIQ-PAQ2PIQ, with ablations on LMBD.","tokens_in":11136,"tokens_out":10551,"duration_ms":104458,"significance":"If the claims are substantiated, the paper makes a useful practical contribution: the idea of exploiting smartphone short-exposure captures as structural guidance for local motion deblurring is well motivated, and the comparison includes retrained baselines with and without the short-exposure image as additional input, which is a fair control. The ablations show each proposed module contributes at least some positive effect, and a code link is provided. I also do not see a circularity problem: the short-exposure image is a genuine extra input rather than a reparametrization of the output. However, the evidence is currently insufficient for the headline claim. All comparisons rely on no-reference metrics, the real-world dataset has no ground-truth sharp image, and the synthetic-to-real transfer of short-exposure images is not validated. The significance is therefore conditional on additional evaluation.","major_comments":[{"comment":"The central comparison is based exclusively on the no-reference metrics NIQE, CLIPIQA+, and MUSIQ-PAQ2PIQ, even though the ReLoBlur synthetic dataset provides ground-truth sharp images. No PSNR, SSIM, LPIPS, or any distortion metric is reported, and no user study is conducted. Since a generative diffusion prior can improve no-reference quality scores while hallucinating content that does not match the scene, the claim that ExpRDiff 'performs favorably against state-of-the-art methods' is not yet established. Please report reference-based metrics on ReLoBlur and a perceptual user study on LMBD, and ideally include a diffusion-based restoration baseline in the comparison.","section":"Section 4.1, Tables 1-2"},{"comment":"The key premise of the method is that the short-exposure image R is a reliable structural reference, but during training R is synthesized by converting sharp images to HSV, reducing brightness, and adding random noise, while at inference R is a real smartphone capture. The manuscript reports no alignment preprocessing for LMBD, no analysis of the distribution shift between synthetic and real R, and no validation that the real R contains the moving object in the same pose. If R is misaligned or has a different noise/color distribution, the masked guided filter in Eq. (4) can copy incorrect structure into blurry regions, so the large NIQE gains in Table 2 could come from the SDXL prior rather than from R. The manuscript repeatedly refers to a supplemental material that is not included in the arXiv submission, so the LMBD capture protocol cannot be checked. Please quantify alignment error or provide a real-pair study, for example by comparing against a deliberately misaligned R or a synthetic R.","section":"Section 4.1 and Eqs. (4)-(6)"},{"comment":"The ablations use only CLIPIQA+ and MUSIQ-PAQ2PIQ on LMBD, and the reported differences are small. For example, in Table 5, 'w/ M' versus 'w/ M^gt' differs by 0.15 in MUSIQ-PAQ2PIQ and 0.004 in CLIPIQA+; in Table 4, 'w/ F & w/o M' versus 'w/ F' differs by 0.44 in MUSIQ-PAQ2PIQ. No standard deviations, confidence intervals, or significance tests are reported, so the claim that each designed module contributes is not quantitatively supported. Please provide multiple runs or significance testing, and include at least one reference-based metric or a user study for the ablations.","section":"Section 5, Tables 3-5"},{"comment":"The training and evaluation of the LMBD model are underspecified. The text states that LMBD contains 10,380 images for training and 209 for evaluation, but it does not describe how ground-truth masks M^gt are obtained for real LMBD images, even though those masks are needed for the detection loss in Eq. (2) and for the 'w/ M^gt' ablation in Table 5. It also does not clearly state whether the Table 2 baselines are trained on ReLoBlur or LMBD and with which hyperparameters. Please clarify the dataset construction and the exact training/evaluation protocol for each table.","section":"Section 4.1 and Section 5"}],"minor_comments":[{"comment":"The sentence after Table 1 says the proposed method achieves 'higher NIQE' but NIQE is a lower-is-better metric; the text should read 'lower NIQE'.","section":"Section 4.2"},{"comment":"Notation is inconsistent: R denotes both the short-exposure image and the folding operation in Eq. (3), and D denotes the detection network in Eq. (1) but is described as the restoration module in the Figure 2 caption. Please disambiguate these symbols.","section":"Section 3.1 and Figure 2"},{"comment":"The Gumbel-Softmax temperature τ, the binarization threshold for M, and the regularization parameter ϵ are never given values, and the phrase 'a threshold based on the default value of the Gumbel-Softmax distribution' is not well-defined. State the exact values and any annealing schedule.","section":"Sections 3.1-3.2"},{"comment":"The sentence 'We employ the loss function [22] to regularize our network' does not state the loss; the objective should be written out so the training procedure is self-contained.","section":"Section 4.1"},{"comment":"The paper claims that the short-exposure guided diffusion model demonstrates its 'effectiveness and efficiency', but no runtime, FLOPs, parameter count, or memory comparison is reported anywhere in the manuscript.","section":"Section 3.3"},{"comment":"There are typos such as '2st patch' in Figures 1 and 6, and the formula for b_k in Eq. (6) is typeset without a visible division by its denominator. Please fix these presentation issues.","section":"Figures and Eq. (6)"}],"recommendation":"major_revision","confidential_remarks":"The paper has a sound architectural idea and fair baseline controls, but the evidence for the headline claim is not yet at the standard expected for a main-track publication. The main risk is the absence of any reference-based metric or user study, and the second risk is the unvalidated domain gap between synthetic and real short-exposure images. I recommend a major revision rather than rejection because the issues are addressable within the scope of the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: ExpRDiff is a coherent engineering contribution that sensibly combines context-based blur detection, masked guided filtering, and short-exposure-conditioned diffusion. The architecture is plausible, the ablations support the importance of each component, and the visual results on the sample images are genuinely less blurry. But the evaluation section undercuts the headline claim in ways that are fixable but need addressing.\n\nWhat's actually new is the integration pattern. Patch-level blur classification with Gumbel-Softmax is not new, but adding context via larger patches to the detector is a sensible tweak. The masked guided filter that only applies the filter in blurred regions is a clean way to keep the background crisp. ExpBFusion, which injects short-exposure and mask features into the SDXL U-Net through cross-attention, is a reasonable conditioning mechanism. The ablations in Tables 3-5 show each module helps, and the effect sizes are nontrivial.\n\nThe soft spots are in the evidence, not the architecture. First, ReLoBlur has ground truth, but the paper reports only no-reference metrics (NIQE, CLIPIQA+, MUSIQ-PAQ2PIQ). PSNR/SSIM are missing. That matters here because diffusion models often trade distortion for perceptual quality; without distortion numbers, we cannot tell how much the method actually preserves content. Second, no diffusion-based baseline is compared, even though the related work cites PASD and DiffBIR. A comparison against at least one of them is necessary to support the claim that the diffusion prior is doing useful work. Third, no error bars or significance tests. The reported gains might be within noise, especially on the small 209-image test set.\n\nThe real-dataset concern raised in the stress test is legitimate but I would phrase it differently. The paper trains on ReLoBlur with synthetic short-exposure frames (HSV brightness reduction + noise) and also collects LMBD, but it never spells out whether the Table 2 model is trained on real LMBD pairs or fine-tuned from ReLoBlur. If it is trained on real pairs, the synthetic generation is less critical; if not, the domain shift is load-bearing. The authors should clarify the training protocol and describe the LMBD capture and alignment procedures. That level of detail is missing.\n\nBottom line: the paper deserves peer review. The problem is practical, the design is sensible, and the flaws are addressable rather than fatal. A serious referee should ask for distortion metrics on ReLoBlur, a diffusion baseline, error bars, and a fuller description of LMBD and the training protocol. I would accept the paper if those are met.\n\nRegards,","headline":"A sensible engineering recipe for local motion deblurring with short-exposure guidance, but the evaluation needs distortion metrics, a diffusion baseline, and a clearer dataset description to support the headline gains.","tokens_in":11678,"tokens_out":5477,"would_cite":false,"duration_ms":48805,"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":"ExpRDiff uses smartphone short-exposure frames to guide a diffusion model in removing local motion blur, reporting better perceptual quality than state-of-the-art deblurring methods on both a synthetic benchmark and a new real-world…","keywords":["local motion deblurring","short-exposure guidance","diffusion model","blur detection","guided image filtering","SDXL","perceptual quality","smartphone photography"],"falsifier":"A concrete test: take real paired long/short-exposure captures from LMBD, deliberately misalign the short-exposure image by a few pixels, and measure the perceptual scores of the restored image. If the scores stay at the reported level, the short-exposure guidance is not actually carrying the result; if they drop sharply, the premise that aligned sharp short-exposure structure drives the improvement is confirmed. A second check: replace the synthetic short-exposure images used in ReLoBlur training with real short-exposure captures and see whether the gain over the synthetic-trained model disappears.","tokens_in":10712,"feed_emoji":"📷","tokens_out":4616,"duration_ms":41898,"temperature":0.7,"pith_summary":"The paper introduces ExpRDiff, a network for removing blur caused by moving objects in photos. It claims that a short-exposure frame of the same scene, which captures the moving object sharply, can be turned into reliable guidance for restoration by first detecting blurry regions, applying a masked guided filter, and then feeding fused short-exposure and blur features into an SDXL diffusion model. The central bet is that this guidance lets a diffusion prior restore realistic detail in heavily blurred areas without smoothing the clear background. If correct, the recipe gives smartphone cameras a practical way to turn a quick capture into a sharp photo of a moving scene.","feed_headline":"Short-exposure frames guide diffusion to erase local motion blur","feed_subtitle":"A masked filter adds sharp short-exposure detail to SDXL, beating prior deblurring methods in perceptual tests.","key_machinery":"The load-bearing object is the short-exposure image itself, treated as a structural prior for the moving objects. ExpRDiff couples three uses of it: a context-based local blur detection module that classifies image patches with a Gumbel-Softmax confidence map; a blurry-aware guided image restoration module that solves a masked guided-filter optimization so that only blurry regions adopt the short-exposure structure while clear background stays as captured; and a short-exposure guided diffusion model whose ExpBFusion module injects mask-weighted short-exposure and blur features into the SDXL U-Net through attention. The mask from the detector also eliminates boundary artifacts left by the guided filter.","core_discovery":"The central claim is that short-exposure-guided diffusion, not merely having the extra reference image, is what makes local motion deblurring work well. Given a long-exposure blurry image and a short-exposure image of the same scene, ExpRDiff first learns a context-based blur detector that marks blurry patches by aggregating patch-level confidence values, then applies a blur-aware masked guided filter that transfers sharp structure from the short-exposure frame only inside the detected blurry regions while leaving clear background untouched. A short-exposure guided diffusion model then refines the result: an ExpBFusion module injects mask-weighted features from both the short-exposure image and the detected blurry regions into the SDXL U-Net through attention, which the paper shows also removes boundary artifacts left by the guided filter. On the ReLoBlur synthetic benchmark and the newly captured LMBD dataset, ExpRDiff reports lower NIQE and higher CLIPIQA+ and MUSIQ-PAQ2PIQ than global deblurring baselines, local deblurring methods, and the same baselines given the short-exposure image as an extra input.","pith_inferences":["The synthesis of short-exposure images by lowering HSV brightness and adding random noise is a cheap proxy for real short-exposure physics; a natural extension is to learn the proxy from real paired captures or to model the exposure-difference transfer function explicitly.","Because the blur mask is patch-level and upsampled with nearest-neighbor interpolation, fine object boundaries are approximated; a boundary-refinement module or instance-aware segmentation could sharpen the mask and further improve edge detail.","The reported metrics (NIQE, CLIPIQA+, MUSIQ-PAQ2PIQ) align with the paper's claim about visual quality, but a side-by-side human preference study would test whether the perceptual gains are noticeable to viewers.","The same guided-fusion recipe may transfer to other local degradations where a short-exposure or differently-exposed frame exists, such as local glare, reflections, or rolling-shutter artifacts."],"forward_implications":["Smartphone burst or multi-camera captures become a practical input: a short-exposure frame of a moving object can serve as guidance, fitting the workflow of cameras that already grab extra frames.","Because the mask keeps clear background untouched, users avoid the over-smoothing that global deblurring applies to static regions.","The diffusion prior can hallucinate realistic texture in heavily blurred regions, so the restored image looks natural even where the short-exposure frame itself is noisy or underexposed.","The approach inherits the computational cost of an SDXL diffusion backbone, so deployment requires a diffusion-capable runtime rather than the lightweight CNN deblurring models it outperforms.","Training on synthesized short-exposure images together with a small real dataset transfers to real-world local blur, as demonstrated on the LMBD dataset."],"supporting_citations":[{"why":"Supplies the ReLoBlur dataset used for training and evaluation, and defines the local-blur benchmark the method must outperform.","marker":"[10]"},{"why":"Provides the guided image filtering formulation that the blurry-aware masked restoration module extends.","marker":"[5]"},{"why":"SDXL is the frozen diffusion backbone that the short-exposure guidance is plugged into.","marker":"[16]"},{"why":"Contributes the diffusion guidance design and the training loss used to regularize the network.","marker":"[22]"},{"why":"LMD-ViT is a state-of-the-art local motion deblurring transformer used as a comparison baseline for the claimed gains.","marker":"[11]"},{"why":"NAFNet is a strong global deblurring baseline that the method must outperform, and it is also retrained with short-exposure input for comparison.","marker":"[3]"},{"why":"Restormer is another strong global baseline, also evaluated with the short-exposure image concatenated as input.","marker":"[26]"}],"fun_headline_variants":["Short-exposure guidance sharpens diffusion-based motion deblurring","Diffusion model uses short-exposure frames to erase local blur","Context-aware blur detection plus guided diffusion beats prior methods","ExpRDiff: short-exposure details guide diffusion for realistic deblurring"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"During training, short-exposure images are simulated by lowering brightness in HSV space and adding random noise to sharp frames (Section 4.1); if real short-exposure captures are misaligned with the long-exposure frame, have a different noise profile, or miss the moving object, the guidance can inject wrong structure and the deblurring gains shrink.","fun_headline_variants_meta":{"raw":{"variants":["Short-exposure guidance sharpens diffusion-based motion deblurring","Diffusion model uses short-exposure frames to erase local blur","Context-aware blur detection plus guided diffusion beats prior methods","ExpRDiff: short-exposure details guide diffusion for realistic deblurring"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000208,"raw_usage":{"total_tokens":1406,"prompt_tokens":952,"completion_tokens":454,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":568,"completion_tokens_details":{"reasoning_tokens":380}},"tokens_in":568,"tokens_out":454,"duration_ms":4634,"temperature":1.0,"reasoning_tokens":380,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T17:13:18.260243+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete test: take real paired long/short-exposure captures from LMBD, deliberately misalign the short-exposure image by a few pixels, and measure the perceptual scores of the restored image. If the scores stay at the reported level, the short-exposure guidance is not actually carrying the result; if they drop sharply, the premise that aligned sharp short-exposure structure drives the improvement is confirmed. A second check: replace the synthetic short-exposure images used in ReLoBlur training with real short-exposure captures and see whether the gain over the synthetic-trained model disappears.","supporting_citations":[{"cited_title":"Real-world deep local motion deblur- ring","cited_arxiv_id":null,"evidence_quote":"Supplies the ReLoBlur dataset used for training and evaluation, and defines the local-blur benchmark the method must outperform."},{"cited_title":"Guided image fil- tering","cited_arxiv_id":null,"evidence_quote":"Provides the guided image filtering formulation that the blurry-aware masked restoration module extends."},{"cited_title":"SDXL: improving latent diffusion mod- els for high-resolution image synthesis","cited_arxiv_id":null,"evidence_quote":"SDXL is the frozen diffusion backbone that the short-exposure guidance is plugged into."},{"cited_title":"Adaptive window prun- ing for efficient local motion deblurring","cited_arxiv_id":null,"evidence_quote":"LMD-ViT is a state-of-the-art local motion deblurring transformer used as a comparison baseline for the claimed gains."},{"cited_title":"Simple baselines for image restoration","cited_arxiv_id":null,"evidence_quote":"NAFNet is a strong global deblurring baseline that the method must outperform, and it is also retrained with short-exposure input for comparison."},{"cited_title":"Restormer: Efficient transformer for high-resolution image restoration","cited_arxiv_id":null,"evidence_quote":"Restormer is another strong global baseline, also evaluated with the short-exposure image concatenated as input."}],"review_version":1}