{"id":"bc84db39-0367-4fc6-a04a-db1150dd8c04","arxiv_id":"2501.12604","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A video diffusion model combined with diffusion posterior sampling reconstructs sharp video frames from a single motion-blurred image by treating the blur as temporal averaging.","lead":"This paper treats a motion-blurred photo as the average of a short video and uses a video diffusion model to reconstruct the sharp frames. If it scales to general scenes, it could turn single-image deblurring into a generative video problem instead of a kernel-estimation problem.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed superiority over baselines is unsupported as stated: baselines are evaluated without training on the test domain while the video diffusion model is trained on it, so the large PSNR gap is not evidence of a better deblurring formulation.","rationale":"The paper proposes a conceptually interesting re-framing of single-image deblurring as a temporal-average inversion problem and implements it with DPS in latent video space. The synthetic CLEVRER results show the mechanism can work when the VDM has seen the exact world and the blur is generated by the assumed average model. However, the quantitative case for 'outperforming existing techniques' rests on a comparison in which baselines are not adapted to the test domain. MPRNet, MTRNN, and Restormer are general-purpose deblurring networks; applying them without retraining to synthetic blur on BAIR/CLEVRER is not a fair test of whether the temporal-averaging formulation is superior. The paper's own caveat that the VDM must have learned the 'given world' substantially narrows the central claim. Additionally, no real motion-blurred image is ever tested; the BAIR experiments synthesize blur by frame averaging, so 'real-world' in the abstract is misleading. These concerns are consistent with the reader's conditional verdict: the conceptual contribution is worthy, but the empirical validation does not currently support the strong claim. A concrete fix is to retrain baselines in-domain and, ideally, evaluate on genuinely motion-blurred video frames using a VDM trained on large-scale real video. Until then, conditional acceptance is appropriate.","tokens_in":6621,"tokens_out":4026,"duration_ms":38939,"concrete_test":"Retrain or fine-tune the three baselines (MPRNet, MTRNN, Restormer) on the same BAIR and CLEVRER training clips used for VDM-MD, synthesizing blurry inputs exactly as in the paper (averaging 10 frames), and then re-evaluate all methods on the same test splits used in Table 2. If the PSNR gap between VDM-MD and the best baseline shrinks to a small margin (e.g., under 2 dB), the claim of superiority is not supported; if the gap remains large, the domain-shift objection is weakened. Report both the gap and whether the baselines converge to reasonable performance on the training distribution.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that VDM-MD outperforms existing single-image deblurring methods on complex motion blur. The empirical support (Table 2) compares VDM-MD, which is trained on 260K clips from BAIR, with MPRNet, MTRNN, and Restormer, which are pre-trained on natural-image deblurring datasets like GoPro and are not fine-tuned or retrained on the BAIR or CLEVRER distribution. The observed gap (e.g., 24.24 dB vs 16.58 dB PSNR on BAIR Main) therefore conflates method quality with domain shift: the baselines are being asked to deblur synthetic frame-average blur in an unfamiliar visual domain, while VDM-MD is evaluated in-distribution. The paper's own Section 3 states the method works 'as long as the image is about the given world,' so the general claim in the abstract and Section 1 that the method 'outperforms existing techniques in deblurring complex motion blur scenarios' is not established by these experiments. This is a load-bearing issue because the entire practical advantage of the approach rests on this empirical comparison.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces VDM-MD, a single-image motion deblurring method that models a blurry image as the temporal average of a short video sequence and reconstructs the sharp video frames using a pre-trained video diffusion model within a Diffusion Posterior Sampling (DPS) loop. The degradation model is the linear frame-averaging operator of Eq. (4), so no explicit blur kernel is estimated. Experiments on synthetic blur generated from CLEVRER and BAIR videos report high PSNR/SSIM values, and a comparison on BAIR shows large gains over MPRNet, MTRNN, and Restormer.","tokens_in":6842,"tokens_out":7871,"duration_ms":77850,"significance":"The conceptual contribution is attractive: video diffusion models encode temporal dynamics that could resolve the N-to-one ambiguity inherent in temporal averaging, and the DPS formulation is standard. If the empirical claims were fully supported, this would open a promising new direction for deblurring. However, the evaluation currently conflates method quality with domain match: the VDM is trained on the same video distribution used to synthesize the blurry inputs, while the baseline deblurring networks are not. The claimed superiority over existing techniques is therefore not yet established, and the 'real-world dataset' description overstates what is tested.","major_comments":[{"comment":"The comparison with MPRNet, MTRNN, and Restormer is not a fair test of the claimed superiority. VDM-MD is trained on 260K BAIR clips, whereas the baselines are pre-trained on other datasets (e.g., GoPro) and are not fine-tuned or retrained on the BAIR distribution. The large PSNR gap (24.24 vs 16.58 dB on BAIR Main) likely reflects the baselines' out-of-distribution inputs rather than the intrinsic advantage of the temporal-averaging formulation. The paper itself notes at the end of Section 4.2 that the baselines 'were never designed to handle this kind of motion-blur scenario.' To support the abstract's claim that the method outperforms existing techniques, the authors should at least fine-tune the baselines on the same synthetic blur distribution, or evaluate on a common benchmark with a matched video prior for the proposed method.","section":"Section 4.2, Table 2"},{"comment":"The phrase 'real-world datasets' is misleading. While BAIR contains real camera recordings, all blurry inputs are synthesized by averaging consecutive frames, so the degradation model is exactly the assumed H(·). No experiment uses genuine motion-blurred images captured with finite exposure, nor does any test use a blur formation model that deviates from frame averaging (e.g., with sensor noise, non-linear response, or rolling shutter). This matters because the paper's motivation is that real blur is non-convolutional and complex; without tests on actual blurred images, the method's practical advantage over existing deblurring methods is not demonstrated.","section":"Abstract and Section 4"},{"comment":"The 'temporal down-sampling' robustness experiment does not test mismatch with a real blur formation model. Both 'smoothly' and 'less smoothly blurred' images are averages of frames from the same video, just at different frame rates; they still obey the temporal-averaging assumption in Eq. (4). The conclusion that 'H(·) may not perfectly match real-motion conditions' is therefore unsupported. A meaningful mismatch test would require a physical motion-blur synthesis, such as integrating scene radiance along a camera or object trajectory over an exposure interval, or a real blurred image benchmark.","section":"Section 4.1, Table 1"}],"minor_comments":[{"comment":"Reference [18] is incorrect for the CLEVRER dataset; it cites Srivastava et al. (2015) on unsupervised video representation learning. The CLEVRER dataset should be cited as Yi et al., 'CLEVRER: Collision Events for Video Representation and Reasoning,' ICLR 2020.","section":"References, [18]"},{"comment":"The terms 'BAIR Main' and 'BAIR Aux1' are used in Table 2 but never defined. Please specify what these subsets are.","section":"Section 4.2, Table 2"},{"comment":"Algorithm 1 uses the update Z_{t-1} = Z'_{t-1} - η_t ∇_{Z_t} ||y - \\hat y_{t-1}||^2 without the 1/σ^2 factor that appears in Eq. (15). Clarify whether η_t absorbs σ^2 or whether the algorithm is missing the noise-variance normalization.","section":"Algorithm 1"},{"comment":"Equation (15) contains a typo: the term '- β (s_{θ*}(Z_t,t) - ...)' should read '- β(t) ( ... )' to match the notation used elsewhere.","section":"Equation (15)"},{"comment":"CLEVRER is described as having 'minimal motion between consecutive frames,' which suggests the blur being removed may be relatively small. Reporting the average amount of motion or the effective blur kernel width would help assess task difficulty.","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The core idea is original and the mathematical framework is sound, but the empirical evidence for the central claim of outperforming existing deblurring methods is undermined by the in-distribution evaluation and the unfair baseline setup. The paper would be publishable after substantial revisions that either (i) fine-tune the baselines, (ii) include an out-of-distribution or real-blur experiment, or (iii) correctly limit the claims to the trained-video-domain setting. The incorrect CLEVRER citation should also be fixed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuinely new way to think about single-image motion blur — treat the blur as a temporal average of a short video and recover that video with a pretrained video diffusion model via DPS. The paper is worth a serious look, but the empirical section overclaims as written, and the headline comparison against baselines is not fair.\n\nThe conceptual contribution is real. Framing blur as H(X) = (1/N) Σ x_n avoids kernel estimation, and using a video diffusion prior to fill in the missing temporal dimension is a natural fit, even if no one had put it together this way. The DPS-in-latent-space implementation is sound; the math is standard and Algorithm 1 is reproducible in principle. The authors also handle a subtlety honestly: a single blurry image has no temporal direction, so the recovered video can run backward in time. They report that rather than hiding it. The robustness test with temporal down-sampling, where the forward model doesn't exactly match the integral, is a good sanity check.\n\nThe soft spots are in the evaluation. The main claim — that VDM-MD outperforms MPRNet, MTRNN, and Restormer — is not supported by Table 2. Those baselines were trained on natural-image deblurring datasets (GoPro etc.) and are tested on synthetic frame-averages from BAIR, while VDM-MD was trained on 260K BAIR clips. Asking a GoPro-trained network to deblur a BAIR robot arm and then crediting the 8 dB gap to the method is a domain-shift artifact, not a method comparison. The authors even say the baselines' failure is 'unsurprising given they were never designed to handle this kind of motion-blur scenario,' which undercuts their own superiority claim. There is also no evaluation on real motion-blurred images; BAIR is real video but the blur is synthetic averaging of consecutive frames. The paper's own Section 3 limits the method to images 'about the given world,' so the abstract's 'outperforms existing techniques' is too broad.\n\nThese problems are fixable. Fine-tune the baselines on the same training data, or test on a held-out natural-image deblurring benchmark with a VDM trained on something like WebVid. Add at least one real-blur qualitative case. Release code and the exact hyperparameters (the free parameters listed in the reader's report are not given in the paper). Then the conceptual claim can be evaluated properly.\n\nWho is this for: people working on diffusion-based inverse problems and motion deblurring will want to know about this framing. It deserves peer review, but the empirical section needs major revision before publication.","headline":"A genuinely novel framing of single-image deblurring as video reconstruction from a temporal average, but the empirical claims need fair baselines and real-blur evaluation before they hold.","tokens_in":7395,"tokens_out":2155,"would_cite":true,"duration_ms":22497,"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":"Motion blur can be undone by treating a blurry image as the temporal average of a sharp video and letting a video diffusion model recover that video.","keywords":["motion deblurring","video diffusion model","diffusion posterior sampling","temporal averaging","latent diffusion","diffusion transformer","single-image deblurring"],"falsifier":"Record a long-exposure frame and a synchronized high-speed burst of the same real scene, then run the method and compare the recovered frames with the burst frames; if the recovered video is no closer to the burst than the sharpest single burst frame is, the temporal-average-plus-video-prior claim is falsified.","tokens_in":6377,"feed_emoji":"🎬","tokens_out":6172,"duration_ms":65988,"temperature":0.7,"pith_summary":"This paper tries to establish that a single motion-blurred image can be turned back into a short sharp video by modeling blur as an average of frames rather than as a spatial convolution. If true, deblurring would no longer depend on estimating a blur kernel, and one photograph would yield a temporally coherent sequence instead of a single static image. The claim is tested in two controlled video worlds, and the recovered frames are reported to be sharper than those from conventional single-image deblurring baselines.","feed_headline":"A video diffusion model can undo motion blur from a single image","feed_subtitle":"Instead of estimating a blur kernel, it recovers a full frame sequence from one blurred photo using a video prior.","key_machinery":"The load-bearing mechanism is the pair formed by the temporal-average degradation operator $H(X)=\\frac{1}{N}\\sum_{n=0}^{N-1}x_n$ and a pre-trained transformer-based video diffusion prior. In each reverse-diffusion step, the latent video estimate is decoded and the gradient of $\\|y-H(D(\\hat{Z}_0(Z_t)))\\|^2$ is subtracted from the sample update, so the diffusion prior determines which motions are plausible while the blurry image decides which motion actually produced the observation.","core_discovery":"The central claim is that a single motion-blurred image contains a short sharp video hidden by frame averaging, and that a pre-trained video diffusion model can reconstruct that video when the blur is treated as the forward model y = H(X) + e with H(X) the average over frames. The method runs Diffusion Posterior Sampling in a VQ-GAN latent space, so no blur kernel is estimated and the same pipeline covers camera shake, object motion, and non-linear trajectories. The output is a deblurred video rather than one image, with the caveat that the recovered motion can be reversed in time when the blurry image carries no directional information.","pith_inferences":["Going beyond the paper, the same temporal-average inversion could be applied to other temporal degradations, such as rolling-shutter distortion or exposure bracketing, whenever the forward operator is differentiable and a video prior exists.","The observed time-reversal ambiguity implies that a single blurry frame cannot fix the direction of motion; adding a weak directional cue, such as a shadow or one extra frame, would be a natural testable extension.","One could relax the training requirement further by synthesizing blurry images from arbitrary video datasets through frame averaging, potentially scaling the method to domains outside the two toy worlds tested here."],"forward_implications":["A single motion-blurred photograph would become a source of video, not just a static image, whenever a suitable video prior is available.","Deblurring would no longer require kernel estimation or explicit motion segmentation, since the temporal-average model accommodates any motion captured by the video prior.","Pre-trained video diffusion models could be repurposed for motion deblurring without retraining on blurry-sharp image pairs.","The reported robustness to frame-rate mismatch suggests that ordinary video datasets with frame averaging may suffice for training, without high-speed camera data."],"supporting_citations":[{"why":"Supplies the temporal-averaging approximation y ≈ (1/N)Σx_n that replaces the convolution model in this method.","marker":"[10]"},{"why":"Provides the Diffusion Posterior Sampling loop that injects the blurry observation into the reverse diffusion process.","marker":"[11]"},{"why":"Supplies the transformer-based video diffusion architecture adapted as the unconditional video prior.","marker":"[16]"},{"why":"Supplies the VQ-GAN latent space used to make video diffusion computationally tractable.","marker":"[17]"},{"why":"Supplies the score-based SDE formulation on which the reverse diffusion update is built.","marker":"[13]"},{"why":"Motivates the premise that large video diffusion models capture realistic motion and can act as world models.","marker":"[15]"},{"why":"Supplies the synthetic CLEVRER world used for controlled training and evaluation.","marker":"[18]"},{"why":"Supplies the BAIR robot-pushing videos used for evaluation with real-camera imagery.","marker":"[21]"}],"fun_headline_variants":["Video diffusion model deblurs single images without blur kernels","One blurred image becomes a sharp video with diffusion","Treat motion blur as frame averaging, let video diffusion recover","No kernel estimation: video diffusion reverses motion blur"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire method rests on the assumption that a motion-blurred photo is the average of a short burst of sharp frames drawn from the same kind of world the video diffusion model was trained on.","fun_headline_variants_meta":{"raw":{"variants":["Video diffusion model deblurs single images without blur kernels","One blurred image becomes a sharp video with diffusion","Treat motion blur as frame averaging, let video diffusion recover","No kernel estimation: video diffusion reverses motion blur"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00051,"raw_usage":{"total_tokens":2409,"prompt_tokens":798,"completion_tokens":1611,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":414,"completion_tokens_details":{"reasoning_tokens":1548}},"tokens_in":414,"tokens_out":1611,"duration_ms":13288,"temperature":1.0,"reasoning_tokens":1548,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T16:59:30.065234+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record a long-exposure frame and a synchronized high-speed burst of the same real scene, then run the method and compare the recovered frames with the burst frames; if the recovered video is no closer to the burst than the sharpest single burst frame is, the temporal-average-plus-video-prior claim is falsified.","supporting_citations":[{"cited_title":"Event-guided multi-patch network with self-supervision for non-uniform motion deblurring,","cited_arxiv_id":null,"evidence_quote":"Supplies the temporal-averaging approximation y ≈ (1/N)Σx_n that replaces the convolution model in this method."},{"cited_title":"Multi-temporal recurrent neural net- works for progressive non-uniform single image deblur- ring with incremental temporal training,","cited_arxiv_id":null,"evidence_quote":"Provides the Diffusion Posterior Sampling loop that injects the blurry observation into the reverse diffusion process."},{"cited_title":"Deblurgan: Blind motion deblurring using conditional adversarial net- works,","cited_arxiv_id":null,"evidence_quote":"Supplies the score-based SDE formulation on which the reverse diffusion update is built."},{"cited_title":"Id-blau: Im- age deblurring by implicit diffusion-based reblurring augmentation,","cited_arxiv_id":null,"evidence_quote":"Motivates the premise that large video diffusion models capture realistic motion and can act as world models."}],"review_version":1}