Pith. sign in

REVIEW 3 major objections 5 minor 53 references

RealVDeblur claims a one-step diffusion framework can deblur real-world video by training on physically grounded synthetic blur, and shows this transfers to unseen benchmarks and downstream 3D reconstruction.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

RealVDeblur trains a one-step video-diffusion deblurrer on a large 3DGS-based synthetic dataset and stabilizes long-video inference with a temporal window mask, improving perceptual quality on real-world benchmarks.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection Solid engineering with a real evaluation problem: two of the four 'video' benchmarks are image datasets, so the temporal consistency claims need to be re-examined. the 3 major comments →

arxiv 2607.20628 v1 pith:YGHKFACF submitted 2026-07-22 cs.CV cs.AI

RealVDeblur: One-Step Diffusion for Generalizable Real-World Video Deblurring

Renbiao Jin , Mingxin Yang , Yutian Chen , Junhao Zhuang , Xin Cai , Mulin Yu , Linning Xu , Wenxian Yu
show 3 more authors
Danping Zou Shi Guo Tianfan Xue
This is my paper
classification cs.CV cs.AI
keywords real-world video deblurringvideo diffusion modelsone-step distillationblur synthesistemporal window mask3D Gaussian Splattingperceptual qualitytemporal consistency
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to solve the generalization problem in real-world video deblurring: models trained on synthetic blur tend to over-smooth and fail on videos they have not seen. RealVDeblur argues that two bottlenecks — unrealistic training data and the lack of a strong prior over sharp video — can be removed together. It builds a large synthetic training set in which blur is generated by physically rendering camera motion and defocus through 3D scenes and by using high-frame-rate videos for object motion, then adapts a pre-trained video diffusion model to restore sharp frames. For practical use, the multi-step diffusion sampler is distilled into one step, and a training-free temporal window mask keeps long-video inference stable with constant memory. If the claim holds, generative video deblurring becomes fast enough and robust enough for mobile imaging and 3D reconstruction pipelines.

Core claim

The paper's central claim is that a single framework can recover sharp, temporally consistent video from real-world blur across unseen datasets by combining a physically grounded blur synthesis pipeline with a pre-trained video diffusion prior. The evidence it points to: disabling temporal compression in the VAE and encoding frames independently preserves per-frame blur information; distilling the 50-step diffusion process into a one-step generator with pixel-space supervision does not sacrifice quality and even improves it; and a training-free temporal window mask, sized to the training length, removes rotary-embedding extrapolation artifacts while keeping memory constant. Trained on about

What carries the argument

The load-bearing pieces are four. (1) A large-scale training set whose blur comes from 3D Gaussian Splatting (3DGS) with sub-frame rendering along perturbed camera trajectories (camera shake), a thin-lens aperture model (defocus), high-frame-rate videos (object motion), and RAW-domain ISP simulation with Poisson–Gaussian noise. (2) Frame-wise VAE encoding: the causal 3D VAE's temporal compression is switched off and each frame is encoded independently, so the diffusion transformer sees per-frame blur rather than blended latents. (3) One-step distillation via distribution matching with additional ℓ1 and LPIPS losses on decoded frames, producing a single-step generator. (4) A training-free Tem

Load-bearing premise

The overarching real-world claim depends on the synthetic blur — 3DGS sub-frame averaging, thin-lens defocus, and high-frame-rate object motion, all pushed through a simulated RAW/ISP pipeline — acting as a faithful enough proxy for blur produced by real cameras.

What would settle it

Train the same architecture and distillation recipe twice, once on the paper's physically grounded synthetic data and once on conventional RGB frame-averaged blur, and evaluate both on real paired benchmarks such as RealBlur; if the physically grounded model does not win on perceptual and temporal metrics, the synthesis realism is not what drives the reported real-world generalization.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • One-step diffusion restoration at roughly 0.15 seconds per frame becomes a practical preprocessing tool for long real-world videos, including mobile capture and 3D reconstruction pipelines.
  • If the physically grounded synthesis transfers as the experiments suggest, large-scale synthetic data can replace hard-to-collect paired real blur training data for this task.
  • The distilled one-step student matching or exceeding a 50-step teacher, thanks to pixel-space supervision, points to distillation-plus-image-losses as a recipe for efficient generative restoration.
  • The Temporal Window Mask offers a general way to use pre-trained video diffusion models beyond their training length without retraining, applicable to other long-video restoration or generation settings.
  • Improved 3D reconstruction and depth estimation after deblurring suggest that deblurring can be decoupled from reconstruction rather than requiring joint end-to-end optimization.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper leaves implicit that the frame-wise encoding trick likely transfers to other video restoration tasks where per-frame degradation varies (e.g., video super-resolution, rain or snow removal); a direct test would be applying the same modification to a compressed-VAE restoration baseline.
  • An editorial inference: the fixed window size may not be optimal for all motion speeds; adaptive windows tied to estimated blur magnitude are a testable extension the paper does not explore.
  • The attribution of the one-step gain to pixel-space supervision is a hypothesis the paper supports with one ablation; a sharper test would vary the ℓ1 and LPIPS weights independently while holding distribution-matching distillation fixed.
  • The component-wise effect of the synthetic data on real-world generalization could be decomposed further than the paper's single ablation: training on camera-motion-only, defocus-only, and object-motion-only subsets and measuring each benchmark would show which physical mechanism drives the transfer.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. RealVDeblur proposes a one-step diffusion-based video deblurring framework built on a pre-trained video diffusion model (Wan2.1). It makes three main contributions: (1) a frame-wise VAE encoding scheme designed to preserve frame-dependent blur variations; (2) a one-step Distribution Matching Distillation (DMD) pipeline with pixel-space losses for efficient inference; and (3) a training-free Temporal Window Mask (TWM) that restricts self-attention to local temporal windows to avoid RoPE extrapolation artifacts and enable constant-memory long-video processing. To support real-world generalization, the paper also introduces OmniBlur, a large-scale synthetic training dataset generated from 3DGS scenes and high-frame-rate videos with RAW-domain ISP simulation. The method is evaluated on BSD, RealBlur, RSBlur, FEVD, and RWBI, with additional downstream 3DGS reconstruction experiments. The authors report strong perceptual quality, temporal consistency, and downstream robustness relative to several video deblurring baselines.

Significance. If the claims hold, RealVDeblur would be a practically valuable contribution: it demonstrates that a large video diffusion prior can be adapted to video deblurring with one-step inference, and that carefully synthesized training data can improve cross-dataset generalization. The paper's ablations quantify the contribution of each component (frame-wise VAE, 3DGS-rendered data, DMD, TWM), and the downstream 3DGS evaluation is a useful step beyond pixel-level metrics. The central generalization claim, however, depends on the validity and interpretation of the benchmark protocol, particularly the temporal metrics reported on image-based datasets. The core methodology is plausible, but the evaluation weaknesses described below prevent the paper from fully supporting its headline claims as written.

major comments (3)
  1. [Sec. 4.1, Tables 1–2] The paper reports tOF (temporal optical flow error) for RealBlur [31] and RSBlur [30], and states in Sec. 4.1 that it extracts 50 and 165 sequences with an average length of 20 frames per sequence from their test sets. However, RealBlur is a dataset of real blurred/sharp image pairs and RSBlur is a synthetic image-deblurring benchmark; neither provides temporal video sequences. It is therefore unclear how a 20-frame sequence or an optical-flow error can be computed for these benchmarks. The claim in Sec. 4.2 that RealVDeblur achieves the 'best tOF on all four benchmarks' is central to the temporal-consistency validation of the TWM. Please specify exactly how these sequences were constructed, or restrict temporal metrics to benchmarks that have actual video frames (BSD, FEVD).
  2. [Sec. 3.3 vs. Sec. 4.1] RSBlur is listed as an 'unseen' real-world benchmark, but OmniBlur's ISP-aware augmentation is explicitly built 'following RSBlur' (Sec. 3.3), and the rendering pipeline described in the supplement closely mirrors RSBlur's design. As a result, RSBlur is not an independent out-of-distribution test for a model trained on OmniBlur; high performance on RSBlur is partly expected and weakly supports the cross-dataset generalization claim. The paper should either exclude RSBlur from the set of 'unseen' benchmarks or provide an analysis quantifying the overlap between the OmniBlur synthesis pipeline and the RSBlur test generation process.
  3. [Sec. S3.1, Table 3] The downstream 3DGS comparison is not fully controlled. For the cascaded pipelines (VRT+3DGS and RealVDeblur+3DGS), camera poses are estimated via COLMAP on the deblurred images, whereas the end-to-end baselines (BAGS, Deblurring-3DGS) are given the poses and point cloud initialization provided by the DeblurNeRF dataset, reconstructed from the original blurry images. Differences in pose quality can affect reconstruction performance independently of deblurring quality. Additionally, the text in Sec. 4.3 states that RealVDeblur 'consistently outperforms' end-to-end alternatives on camera-motion blur, but Table 3 shows PSNR 25.41 vs. 25.67 for Deblurring-3DGS; the claim holds only for SSIM and LPIPS. Please clarify the experimental protocol and temper the claim accordingly.
minor comments (5)
  1. [Sec. 4.1] The metrics list mentions tOF but does not cite a specific definition or implementation. Please provide a reference and describe how tOF is computed over each benchmark, especially for datasets without ground-truth flow.
  2. [Tables 1–4] No error bars or statistical significance tests are reported. Some improvements are small (e.g., PSNR 28.76 vs. 28.02 in Table 4), and a single run may not be stable. Please report variance or significance over multiple runs or seeds.
  3. [References] VRT is cited as both [14] and [15] with identical title/authors. A single reference should be used. Also, CLIP-IQA in Table 2 is not introduced in the metrics paragraph; please add a citation and definition.
  4. [Sec. 3.2, Eq. (3)] The TWM notation [t−W/2, t+W/2] is ambiguous for even W and for frames near sequence boundaries. Please specify the exact masking rule, including how the window is clipped or padded at the start/end of a video.
  5. [Sec. S1] The condition injection module is described as a 3D convolution with kernel 1×2×2 and stride 1×2×2. Please clarify whether the temporal dimension is preserved and how the 16-channel input becomes 1536 channels; the current one-line description is difficult to reproduce.

Circularity Check

1 steps flagged

Training synthesis is partly copied from the RSBlur test recipe, so RSBlur is an in-distribution benchmark; the central method and other benchmarks remain independent.

specific steps
  1. other [Sec. 3.3 (ISP-Aware Augmentation and Data Composition) and Sec. 4.1 (Benchmarks), Table 1 RSBlur column]
    "Following RSBlur [30], we inject Poisson–Gaussian noise in the RAW domain and simulate the ISP (mosaicing, white balance, demosaicing) to produce realistic training pairs. ... we conduct evaluations on BSD [49], RealBlur [31], RSBlur [30], FEVD [8], and RWBI [46]."

    The OmniBlur training distribution is explicitly built by copying RSBlur's degradation recipe (RAW-domain noise plus ISP simulation), and RSBlur is then used as a held-out 'real-world benchmark' to demonstrate generalization. As a result, the RSBlur row in Table 1 is partially in-distribution with the training-time synthesis by construction: strong RSBlur results are expected from the training-data design rather than from transfer to an unseen degradation. This is a construction-level overlap, not an equation-level reduction, and it does not by itself invalidate the method: BSD, RealBlur, FEVD, and RWBI remain independent validation sets, so the central generalization claim still has independent support.

full rationale

The central derivation chain is self-contained: a pre-trained Wan2.1 DiT is adapted with frame-wise VAE encoding, LoRA conditioning, one-step DMD distillation with pixel-space L1/LPIPS supervision, and a training-free Temporal Window Mask. No fitted parameter is renamed as a prediction, and no equation-level self-definition appears; TWM's W=20 is an explicit design choice that matches the training sequence length, not a fitted result. The only overlapping self-citation (Anyrecon, ref. [1]) appears in a related-work sentence about windowed processing and is not load-bearing. The RSBlur overlap is the one substantive circularity-adjacent concern: Sec. 3.3 says OmniBlur follows RSBlur's ISP/noise simulation, and Sec. 4.1 lists RSBlur as a test benchmark, so RSBlur is not fully 'unseen.' I do not count the reported tOF/sequence-length numbers for RealBlur and RSBlur as circularity; that is an evaluation-validity concern, since the cited datasets are image benchmarks without temporally ordered 20-frame sequences. The supplementary limitations (text distortion, extreme-blur high-frequency loss) are candid and do not indicate circularity. Overall, the method and most benchmarks stand independently, but the RSBlur evaluation is partially compromised by construction, giving a moderate score of 4.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 0 invented entities

No new physical entities are introduced. The empirical claims rest on several hand-set hyperparameters and domain assumptions about synthetic-data fidelity and reuse of a pre-trained video diffusion model. The OmniBlur dataset is a construction, not an invented entity.

free parameters (7)
  • Temporal Window Mask size W = 20
    Set to match training sequence length; ablations show W=5 worse and W=20 best. Hand-set, central to long-video stability.
  • DMD pixel-space loss weights lambda_L1, lambda_LPIPS = 2, 2
    Weights for L1 and LPIPS supervision during one-step distillation; no ablation reported, affects student quality.
  • LoRA rank = 64
    Parameter-efficient adaptation capacity; not swept; impacts the strength of task-tuning.
  • Training sequence lengths and resolutions = 10 frames at 1280x720; 20 frames at 960x544
    Mixed-resolution training choice; the 20-frame length anchors TWM window size.
  • Blur rendering sub-frame/aperture counts = K=121, N=32, compound K=61 x N=16
    Hand-set trade-offs between physical fidelity and rendering cost in OmniBlur synthesis.
  • OmniBlur data composition ratios = 40% camera, 40% defocus, 20% compound; ~53k 3DGS clips + ~3k HFR clips
    Data mix chosen by hand; directly affects generalization across blur types.
  • Fine-tuning schedule = LR 5e-5, 5000 iterations, batch size 32
    Standard hyperparameters; no sweep reported.
axioms (6)
  • domain assumption Pre-trained Wan2.1 VDM provides a useful generative prior for sharp video textures.
    Sec 3.1 builds on Wan2.1; if the prior does not transfer to deblurring, the method fails.
  • domain assumption Wan2.1 DiT accepts frame-wise latents without distribution mismatch because images can be treated as single-frame videos.
    Sec 3.1 'Frame-wise Latent Encoding'; no quantitative latent-distribution check is provided.
  • domain assumption OmniBlur synthetic blur is a faithful proxy for real camera-motion, object-motion, and defocus blur.
    Sec 3.3; the central real-world generalization premise.
  • domain assumption A local temporal window of 20 frames is sufficient for deblurring; global context is unnecessary.
    Sec 3.2 TWM; supported by ablation but not by a principled analysis.
  • domain assumption No-reference metrics (MUSIQ, NIQE, CLIP-IQA) and LPIPS/FID adequately capture perceptual quality and semantic fidelity.
    Sec 4.1; used as the basis for the main claims.
  • standard math Flow-matching objective and DMD score-matching are valid for this latent space.
    Eq. (1) and Sec 3.2 inherit from Lipman et al. and Yin et al.; no new proof is given.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of RealVDeblur: One-Step Diffusion for Generalizable Real-World Video Deblurring." pith.science (2026). https://pith.science/paper/YGHKFACF

@misc{pith2026260720628,
  author       = {Pith},
  title        = {Pith review of: RealVDeblur: One-Step Diffusion for Generalizable Real-World Video Deblurring},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YGHKFACF}},
  note         = {Machine review of arXiv:2607.20628}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Real-world video deblurring remains challenging due to diverse motion patterns, complex degradations, and the scarcity of realistic training data, yet robust restoration is critical for downstream pipelines such as mobile imaging and 3D reconstruction. This work presents \textbf{RealVDeblur}, an efficient generative framework designed to improve in-the-wild robustness under diverse real capture conditions. First, a large-scale, physically grounded blur synthesis pipeline is constructed from scene-level 3D Gaussian Splatting (3DGS) assets and high-frame-rate videos, providing realistic training data covering both camera-induced and object-motion blur. Second, a video diffusion prior is leveraged for restoration; to better accommodate frame-dependent blur variations, temporal compression in the VAE is disabled and a frame-wise encoding scheme is adopted. For practical deployment on long videos, multi-step diffusion sampling is distilled into an efficient one-step generator, and a training-free Temporal Window Mask stabilizes inference beyond the training horizon with constant memory usage. Extensive experiments on diverse real-world benchmarks demonstrate strong perceptual quality, semantic fidelity, and temporal consistency on unseen videos, as well as improved robustness in downstream 3D reconstruction under severe motion blur. Project page: https://rbjin.github.io/RealVDeblur

Figures

Figures reproduced from arXiv: 2607.20628 by Danping Zou, Junhao Zhuang, Linning Xu, Mingxin Yang, Mulin Yu, Renbiao Jin, Shi Guo, Tianfan Xue, Wenxian Yu, Xin Cai, Yutian Chen.

Figure 1
Figure 1. Figure 1: Visual comparison of our proposed RealVDeblur against state-of-the-art video deblurring methods in real-world scenarios. While traditional regression-based methods struggle with in-the-wild complex motion blur and produce over-smoothed re￾sults, our generative framework faithfully recovers sharp structures and high-frequency details. Video deblurring methods have progressed from explicit motion alignment [… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the RealVDeblur framework. Our approach repurposes a pre￾trained Video Diffusion Model (VDM) for video deblurring with three key components: (1) a frame-wise VAE encoding scheme that preserves frame-dependent blur variations, (2) one-step inference via Distribution Matching Distillation (DMD) with pixel-space supervision, and (3) a training-free Temporal Window Mask that restricts self-attentio… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison on multiple real-world evaluation datasets. Under chal￾lenging conditions including severe motion blur, low light, and overexposure, baseline models typically yield heavily smoothed outputs or structural distortions. In contrast, RealVDeblur(Ours) effectively recovers sharp, high-frequency details while maintain￾ing geometric fidelity. GT denotes ground truth [PITH_FULL_IMAGE:figure… view at source ↗
Figure 4
Figure 4. Figure 4: Visual evaluation of 3DGS reconstruction under real-world camera-motion blur (first row) and defocus blur (second row). Compared with end-to-end methods (BAGS, Deblurring-3DGS) and the cascaded baseline (VRT + 3DGS), using RealVDeblur as a preprocessing step substantially improves geometric accuracy and texture sharpness of the reconstructed scenes. One-Step Distillation (DMD). Counter-intuitively, the dis… view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of Causal 3D VAE vs. Frame-wise VAE. In sequences with rapid motion (e.g., Frames 27-32), the Causal 3D VAE exhibits significant artifacts and temporal incoherence, while the Frame-wise VAE maintains sharp details and high reconstruction fidelity. Input w/o TWM w/ TWM (W=5) w/ TWM (W=20) [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative ablation of the Temporal Window Mask on a 150-frame video across three attention configurations: global attention (w/o TWM), a narrow window (W = 5), and our default window (W = 20). Global attention causes severe artifacts due to RoPE extrapolation; a narrow window avoids extrapolation but limits tempo￾ral aggregation; our default window achieves the best balance between stability and restorat… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

53 extracted references · 8 linked inside Pith

  1. [1]

    arXiv preprint arXiv:2604.19747 (2026) 4

    Chen, Y., Guo, S., Jin, R., Yang, T., Cai, X., Luo, Y., Yang, M., Yu, M., Xu, L., Xue, T.: Anyrecon: Arbitrary-view 3d reconstruction with video diffusion model. arXiv preprint arXiv:2604.19747 (2026) 4

    Pith/arXiv arXiv 2026

  2. [2]

    ACM Transactions on Graphics (TOG)39(4), 75–1 (2020) 9

    Chu, M., Xie, Y., Mayer, J., Leal-Taixé, L., Thuerey, N.: Learning temporal co- herence via self-supervision for gan-based video generation. ACM Transactions on Graphics (TOG)39(4), 75–1 (2020) 9

    2020

  3. [3]

    Advances in neural information processing systems30(2017) 9

    Heusel,M.,Ramsauer,H.,Unterthiner,T.,Nessler,B.,Hochreiter,S.:Ganstrained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems30(2017) 9

    2017

  4. [4]

    Iclr1(2), 3 (2022) 4

    Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. Iclr1(2), 3 (2022) 4

    2022

  5. [5]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Ke, J., Wang, Q., Wang, Y., Milanfar, P., Yang, F.: Musiq: Multi-scale image quality transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5148–5157 (2021) 9

    2021

  6. [6]

    ACM Trans

    Kerbl, B., Kopanas, G., Leimkühler, T., Drettakis, G., et al.: 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph.42(4), 139–1 (2023) 7, 11

    2023

  7. [7]

    Advances in Neural Information Processing Systems37, 80965–80986 (2024) 21

    Kheradmand, S., Rebain, D., Sharma, G., Sun, W., Tseng, Y.C., Isack, H., Kar, A., Tagliasacchi, A., Yi, K.M.: 3d gaussian splatting as markov chain monte carlo. Advances in Neural Information Processing Systems37, 80965–80986 (2024) 21

    2024

  8. [8]

    Kim, T., Cho, H., Yoon, K.J.: Frequency-aware event-based video deblurring for real-worldmotionblur.In:ProceedingsoftheIEEE/CVFConferenceonComputer Vision and Pattern Recognition. pp. 24966–24976 (2024) 4, 8, 9

    2024

  9. [9]

    In: The Thirty-ninth Annual Conference on Neural Information Processing Systems (2025) 4

    Kong, L., Zou, D., Wang, F.L., Ren, J., Wu, X., Dong, J., Pan, J., et al.: Deblurdiff: Real-word image deblurring with generative diffusion models. In: The Thirty-ninth Annual Conference on Neural Information Processing Systems (2025) 4

    2025

  10. [10]

    In: European Conference on Computer Vision

    Lee, B., Lee, H., Sun, X., Ali, U., Park, E.: Deblurring 3d gaussian splatting. In: European Conference on Computer Vision. pp. 127–143. Springer (2024) 11, 22

    2024

  11. [11]

    Lee, D., Park, J., Lee, K.M.: Gs-blur: A 3d scene-based dataset for realistic image deblurring.AdvancesinNeuralInformationProcessingSystems37,125394–125415 (2024) 2, 4, 8

    2024

  12. [12]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Li,D.,Shi,X.,Zhang,Y.,Cheung,K.C.,See,S.,Wang,X.,Qin,H.,Li,H.:Asimple baseline for video restoration with grouped spatial-temporal shift. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9822–9832 (2023) 9, 10

    2023

  13. [13]

    In: International Conference on Multimedia Modeling

    Li, Y., Liu, Z., Monno, Y., Okutomi, M.: Tdm: Temporally-consistent diffusion model for all-in-one real-world video restoration. In: International Conference on Multimedia Modeling. pp. 155–169. Springer (2025) 4

    2025

  14. [14]

    IEEE Transactions on Image Processing 33, 2171–2182 (2024) 2, 4, 11, 21

    Liang, J., Cao, J., Fan, Y., Zhang, K., Ranjan, R., Li, Y., Timofte, R., Van Gool, L.: Vrt: A video restoration transformer. IEEE Transactions on Image Processing 33, 2171–2182 (2024) 2, 4, 11, 21

    2024

  15. [15]

    IEEE Transactions on Image Processing 33, 2171–2182 (2024) 9, 10, 11

    Liang, J., Cao, J., Fan, Y., Zhang, K., Ranjan, R., Li, Y., Timofte, R., Van Gool, L.: Vrt: A video restoration transformer. IEEE Transactions on Image Processing 33, 2171–2182 (2024) 9, 10, 11

    2024

  16. [16]

    Advances in Neural Information Processing Systems35, 378– 393 (2022) 2, 4, 9, 10 16 R

    Liang, J., Fan, Y., Xiang, X., Ranjan, R., Ilg, E., Green, S., Cao, J., Zhang, K., Timofte, R., Gool, L.V.: Recurrent video restoration transformer with guided de- formable attention. Advances in Neural Information Processing Systems35, 378– 393 (2022) 2, 4, 9, 10 16 R. Jin et al

    2022

  17. [17]

    arXiv preprint arXiv:2511.10647 (2025) 22

    Lin, H., Chen, S., Liew, J., Chen, D.Y., Li, Z., Shi, G., Feng, J., Kang, B.: Depth anything 3: Recovering the visual space from any views. arXiv preprint arXiv:2511.10647 (2025) 22

    Pith/arXiv arXiv 2025

  18. [18]

    In: European conference on computer vision

    Lin, X., He, J., Chen, Z., Lyu, Z., Dai, B., Yu, F., Qiao, Y., Ouyang, W., Dong, C.: Diffbir: Toward blind image restoration with generative diffusion prior. In: European conference on computer vision. pp. 430–448. Springer (2024) 4

    2024

  19. [19]

    arXiv preprint arXiv:2210.02747 (2022) 5

    Lipman, Y., Chen, R.T., Ben-Hamu, H., Nickel, M., Le, M.: Flow matching for generative modeling. arXiv preprint arXiv:2210.02747 (2022) 5

    Pith/arXiv arXiv 2022

  20. [20]

    arXiv preprint arXiv:2412.00773 (2024) 3, 4

    Long, H., Wang, Y., Wang, W.: Divd: Deblurring with improved video diffusion model. arXiv preprint arXiv:2412.00773 (2024) 3, 4

    Pith/arXiv arXiv 2024

  21. [21]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Ma, L., Li, X., Liao, J., Zhang, Q., Wang, X., Wang, J., Sander, P.V.: Deblur- nerf: Neural radiance fields from blurry images. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12861–12870 (2022) 11, 21

    2022

  22. [22]

    completely blind

    Mittal, A., Soundararajan, R., Bovik, A.C.: Making a “completely blind” image quality analyzer. IEEE Signal processing letters20(3), 209–212 (2012) 9

    2012

  23. [23]

    In: CVPR Workshops (June 2019) 3, 8, 20

    Nah, S., Baik, S., Hong, S., Moon, G., Son, S., Timofte, R., Lee, K.M.: Ntire 2019 challenge on video deblurring and super-resolution: Dataset and study. In: CVPR Workshops (June 2019) 3, 8, 20

    2019

  24. [24]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3883–3891 (2017) 3, 8, 20

    2017

  25. [25]

    arXiv preprint arXiv:2506.19445 (2025) 3, 8, 20

    Noki, M.M.H., Mahmud, S.M., Majumder, P.S., Al Radi, A.M., Ali, M.H., Khan, M.M.: Deblurring in the wild: A real-world dataset from smartphone high-speed videos. arXiv preprint arXiv:2506.19445 (2025) 3, 8, 20

    Pith/arXiv arXiv 2025

  26. [26]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Pan,J.,Bai,H.,Tang,J.:Cascadeddeepvideodeblurringusingtemporalsharpness prior. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3043–3051 (2020) 2, 4

    2020

  27. [27]

    Peebles,W.,Xie,S.:Scalablediffusionmodelswithtransformers.In:Proceedingsof the IEEE/CVF international conference on computer vision. pp. 4195–4205 (2023) 4, 5

    2023

  28. [28]

    In: European Conference on Computer Vision

    Peng, C., Tang, Y., Zhou, Y., Wang, N., Liu, X., Li, D., Chellappa, R.: Bags: Blur agnostic gaussian splatting through multi-scale kernel modeling. In: European Conference on Computer Vision. pp. 293–310. Springer (2024) 11, 22

    2024

  29. [29]

    In: European Conference on Computer Vision

    Rao, C., Li, G., Lan, Z., Sun, J., Luan, J., Xing, W., Zhao, L., Lin, H., Dong, J., Zhang, D.: Rethinking video deblurring with wavelet-aware dynamic transformer and diffusion model. In: European Conference on Computer Vision. pp. 421–437. Springer (2024) 3, 4

    2024

  30. [30]

    In: European conference on computer vision

    Rim, J., Kim, G., Kim, J., Lee, J., Lee, S., Cho, S.: Realistic blur synthesis for learning image deblurring. In: European conference on computer vision. pp. 487–

  31. [31]

    In: European conference on computer vision

    Rim, J., Lee, H., Won, J., Cho, S.: Real-world blur dataset for learning and bench- marking deblurring algorithms. In: European conference on computer vision. pp. 184–201. Springer (2020) 4, 8, 9

    2020

  32. [32]

    arXiv preprint arXiv:2202.00512 (2022) 4

    Salimans, T., Ho, J.: Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512 (2022) 4

    Pith/arXiv arXiv 2022

  33. [33]

    ACM Transactions on Graphics (TOG)40(5), 1–18 (2021) 2, 4 RealVDeblur 17

    Son, H., Lee, J., Lee, J., Cho, S., Lee, S.: Recurrent video deblurring with blur- invariant motion estimation and pixel volumes. ACM Transactions on Graphics (TOG)40(5), 1–18 (2021) 2, 4 RealVDeblur 17

    2021

  34. [34]

    co / datasets / spatialverse/InteriorGS(2025) 3, 8, 19

    SpatialVerse Research Team, M.T.I.: Interiorgs: A 3d gaussian splatting dataset of semantically labeled indoor scenes.https : / / huggingface . co / datasets / spatialverse/InteriorGS(2025) 3, 8, 19

    2025

  35. [35]

    Neurocomputing568, 127063 (2024) 7

    Su, J., Ahmed, M., Lu, Y., Pan, S., Bo, W., Liu, Y.: Roformer: Enhanced trans- former with rotary position embedding. Neurocomputing568, 127063 (2024) 7

    2024

  36. [36]

    In: Proceedings of the IEEE conference on com- puter vision and pattern recognition

    Su, S., Delbracio, M., Wang, J., Sapiro, G., Heidrich, W., Wang, O.: Deep video deblurring for hand-held cameras. In: Proceedings of the IEEE conference on com- puter vision and pattern recognition. pp. 1279–1288 (2017) 8

    2017

  37. [37]

    arXiv preprint arXiv:2503.20314 (2025) 4, 5, 6

    Wan, T., Wang, A., Ai, B., Wen, B., Mao, C., Xie, C.W., Chen, D., Yu, F., Zhao, H., Yang, J., et al.: Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314 (2025) 4, 5, 6

    Pith/arXiv arXiv 2025

  38. [38]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops

    Wang, X., Chan, K.C., Yu, K., Dong, C., Change Loy, C.: Edvr: Video restora- tion with enhanced deformable convolutional networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 0–0 (2019) 2, 4

    2019

  39. [39]

    Wei, X., Liu, X., Zang, Y., Dong, X., Zhang, P., Cao, Y., Tong, J., Duan, H., Guo, Q., Wang, J., et al.: Videorope: What makes for good video rotary position embedding? arXiv preprint arXiv:2502.05173 (2025) 7

    Pith/arXiv arXiv 2025

  40. [40]

    In: European Conference on Computer Vision

    Yang,Q.,Chen,H.,Zhang,Y.,Xia,M.,Cun,X.,Su,Z.,Shan,Y.:Noisecalibration: Plug-and-play content-preserving video enhancement using pre-trained video diffu- sion models. In: European Conference on Computer Vision. pp. 307–326. Springer (2024) 4

    2024

  41. [41]

    Journal of Machine Learning Research26(34), 1–17 (2025) 21

    Ye, V., Li, R., Kerr, J., Turkulainen, M., Yi, B., Pan, Z., Seiskari, O., Ye, J., Hu, J., Tancik, M., et al.: gsplat: An open-source library for gaussian splatting. Journal of Machine Learning Research26(34), 1–17 (2025) 21

    2025

  42. [42]

    arXiv preprint arXiv:2407.01519 (2024) 4

    Yeh, C.H., Shiu, H.S., Lin, C.Y., Wang, Z., Hsiao, C.W., Chen, T.H., Liu, Y.L.: Diffir2vr-zero: Zero-shot video restoration with diffusion-based image restoration models. arXiv preprint arXiv:2407.01519 (2024) 4

    arXiv 2024

  43. [43]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Yeshwanth, C., Liu, Y.C., Nießner, M., Dai, A.: Scannet++: A high-fidelity dataset of 3d indoor scenes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12–22 (2023) 3, 8, 19, 22

    2023

  44. [44]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Yin, T., Gharbi, M., Zhang, R., Shechtman, E., Durand, F., Freeman, W.T., Park, T.: One-step diffusion with distribution matching distillation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6613– 6623 (2024) 4, 7

    2024

  45. [45]

    In: Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition

    Zhang, H., Xie, H., Yao, H.: Blur-aware spatio-temporal sparse transformer for video deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition. pp. 2673–2681 (2024) 2, 4, 9, 10

    2024

  46. [46]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Zhang, K., Luo, W., Zhong, Y., Ma, L., Stenger, B., Liu, W., Li, H.: Deblurring by realistic blurring. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2737–2746 (2020) 8, 10

    2020

  47. [47]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586–595 (2018) 7, 9

    2018

  48. [48]

    In: European conference on computer vision

    Zhong,Z.,Gao,Y.,Zheng,Y.,Zheng,B.:Efficientspatio-temporalrecurrentneural network for video deblurring. In: European conference on computer vision. pp. 191–

  49. [49]

    International Journal of Computer Vision131(1), 284–301 (2023) 4, 8, 9 18 R

    Zhong, Z., Gao, Y., Zheng, Y., Zheng, B., Sato, I.: Real-world video deblurring: A benchmark dataset and an efficient recurrent neural network. International Journal of Computer Vision131(1), 284–301 (2023) 4, 8, 9 18 R. Jin et al

    2023

  50. [50]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Zhou, S., Zhang, J., Pan, J., Xie, H., Zuo, W., Ren, J.: Spatio-temporal filter adap- tive network for video deblurring. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2482–2491 (2019) 2, 4

    2019

  51. [51]

    In: Proceedings of the AAAI conference on artificial intelligence

    Zhu, C., Dong, H., Pan, J., Liang, B., Huang, Y., Fu, L., Wang, F.: Deep recurrent neural network with multi-scale bi-directional propagation for video deblurring. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 3598–3607 (2022) 9, 10 RealVDeblur: Supplementary Material S1 Implementation Details Frame-wise VAE Encoding.Wan-...

    2022

  52. [207]

    Springer (2020) 4, 9, 10

    2020

  53. [503]

    Springer (2022) 2, 4, 8, 9, 21

    2022

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.