PickMoment: Continuous-Time Single-Image-to-Video via Learning Deblurring and Blur-to-Video

Junseong Shin; Hyeonsu Jo; Daehyun Kim; Tae Hyun Kim

cs.CVOct 1, 2026

PickMoment: Continuous-Time Single-Image-to-Video via Learning Deblurring and Blur-to-Video

Authors: Junseong Shin, Hyeonsu Jo, Daehyun Kim, Tae Hyun Kim

Organizations: Department of Artificial Intelligence, Hanyang University · Department of Intelligence Convergence, Hanyang University · Department of Computer Science, Hanyang University

Abstract

Motion blur arises from the temporal integration of a continuous sharp signal over a finite exposure window, yet existing learning-based methods sidestep this physical model and predict only the sharp signal itself: most single-image deblurring methods recover a single frame at the exposure center, while blur-to-video methods predict a fixed set of frames. We introduce PickMoment, a continuous-time reformulation that directly learns the interval-mean blur over arbitrary sub-intervals of the exposure with a single deterministic model. Drawing an analogy to MeanFlow's average-velocity formulation, we train the model with three supervisions derived from the blur integral: an empirical reconstruction loss from available subframes, an additivity loss that enforces self-consistency across overlapping sub-intervals, and a sharp-frame loss anchored at the zero-interval limit. A single trained model unifies single-image deblurring, blur-to-video generation, and continuous-time pick-a-moment recovery as different queries to the same network, with no separate training for each task. Our PickMoment achieves state-of-the-art performance among generative-based deblurring methods on GoPro and HIDE while competitive against restoration-based methods on RealBlur, and the highest per-frame fidelity on GoPro-7 blur-to-video, all in a single forward pass without iterative sampling.

Figures & tables

Figure 1 : Overview of the proposed PickMoment framework. Given a single motion-blurred image, our method unifies three tasks: (a) Image deblurring, predicting the center sharp frame of the blur exposure. (b) Blur-to-video generation predicts an $n$ -frame sharp video by uniformly sampling the normalized exposure trajectory at $M$ time points with interval $\Delta t=1/(M-1)$ ; the figure illustrates an 11-frame example. (c) Arbitrary sharp-frame prediction enables the recovery of sharp frames at arbitrary time points within the blur exposure, thereby enabling continuous-time video generation.

Figure 2 : Conceptual illustration of the training strategy of PickMoment.

		Restoration-based				Generative-based
Dataset	Metric	Restormer [ 45 ]	HI-Diff [ 5 ]	UFPNet [ 8 ]	AdaRevD [ 27 ]	DiffBIR [ 20 ]	OSEDiff [ 43 ]	Diff-Plugin [ 24 ]	FideDiff [ 22 ]	Ours
GoPro	PSNR $\uparrow$	32.92	33.33	34.09	34.60	26.15	24.34	22.88	28.79	29.43
	SSIM $\uparrow$	0.9611	0.9642	0.9686	0.9716	0.8377	0.8228	0.7798	0.9148	0.9253
	LPIPS $\downarrow$	0.0841	0.0799	0.0764	0.0712	0.2366	0.1738	0.2332	0.0831	0.0802
	DISTS $\downarrow$	0.0724	0.0710	0.0666	0.0672	0.1460	0.0834	0.1166	0.0525	0.0429
HIDE	PSNR $\uparrow$	31.22	31.46	31.74	32.35	25.94	23.20	21.94	27.28	27.52

Table 1 : Quantitative comparison for single-image deblurring. Bold indicates the best result and underline indicates the second-best result among generative-based methods.

Figure 3 : Qualitative comparison for single-image deblurring on GoPro, HIDE, and RealBlur-J (top to bottom). Best viewed zoomed in.

Table 5

Figure 4 : Qualitative comparison for blur-to-video on GoPro-7. Best viewed zoomed in.

	Deblurring				Blur-to-video
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	DISTS $\downarrow$	PSNR p $\uparrow$	SSIM p $\uparrow$	LPIPS p $\downarrow$	FVD $\downarrow$	EPE $\downarrow$
$\mathbf{B\rightarrow S}$	27.72	0.8944	0.1595	0.1149	29.16	0.8894	0.0292	59.51	1.15
$\mathbf{B\rightarrow\{S(\tau_{i})\}_{i=0}^{N-1}}$	27.76	0.8945	0.1604	0.1166	29.26	0.8933	0.0278	58.76	1.13
$\mathbf{B\rightarrow B}$ (ours)	28.45	0.9080	0.1455	0.1043	29.78	0.9034	0.0271	59.29	1.02

Table 4 : Advantage of interval prediction on GoPro.

\mathbf{B\rightarrow B}

denotes our proposed method.

Figure 5 : Loss-term ablation visualization on GoPro-7 (frame 6 of 7). Best viewed zoomed in.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Backbone	FLUX.2-klein-4B (transformer only fine-tuned, VAE frozen)
Trainable params	$\sim$ 4B (full transformer)
Optimizer	Adam
Total iterations	100K
LR schedule	Linear warmup to $10^{-4}$ over 2K, then cosine-annealing to $10^{-5}$
Mixed precision	bf16
Gradient accumulation	2 steps

Appendix

Table 6: Training hyperparameters.

PSNR (dB)	Description	Mean	Median	Std.	10th pct.	90th pct.
$\mathcal{D}(z_{s,t})$ vs. $B_{s,t}$	(VAE round-trip only)	42.70	43.12	2.32	39.27	45.24
$\mathcal{D}(z_{\text{pred}})$ vs. $\mathcal{D}(z_{s,t})$	(pure additivity)	39.82	39.38	4.49	34.36	45.94
$\mathcal{D}(z_{\text{pred}})$ vs. $B_{s,t}$	(combined)	37.40	37.26	2.80	33.83	41.03

Appendix

Table 7: Latent-space additivity validity. PSNR distributions over

N=2{,}983

random

(s,t,m)

triples drawn from

11

GoPro test sequences (

512\!\times\!512

crops). The pure additivity error

\mathcal{D}(z_{\text{pred}})

vs.

\mathcal{D}(z_{s,t})

is comparable to the VAE round-trip floor, indicating that

\mathcal{E}

is approximately linear over the blur-image manifold.

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	DISTS $\downarrow$
Jin et al. [ 15 ]	29.44	0.9237	0.1345	0.0982
MotionETR [ 48 ]	28.60	0.9415	0.0509	0.0495
Blur2Vid [ 37 ]	30.66	0.9528	0.0363	0.0380
Ours	33.22	0.9664	0.0291	0.0298

Appendix

Table 8: Center frame quality on the GoPro-7 split. 1,744 clips at

640\times 360

. Frame index 3 of the recovered 7-frame sequence is evaluated against the corresponding ground-truth sharp frame using full-image metrics.

Dataset	Metric	$\mathcal{L}_{\text{recon}}$	$\mathcal{L}_{\text{recon}}+\mathcal{L}_{\text{add}}$	$\mathcal{L}_{\text{recon}}+\mathcal{L}_{\text{sharp}}$	Full
GoPro	PSNR $\uparrow$	28.45	28.56	29.24	29.43
	SSIM $\uparrow$	0.9080	0.9096	0.9236	0.9253
	LPIPS $\downarrow$	0.1455	0.1422	0.0787	0.0802
	DISTS $\downarrow$	0.1043	0.1020	0.0420	0.0429
HIDE	PSNR $\uparrow$	26.73	26.79	27.49	27.52
	SSIM $\uparrow$	0.8668	0.8687	0.8894	0.8891

Appendix

Table 9: Cross-dataset loss-term ablation on single-image deblurring metrics. Extension of Table. 5 .

Figure 6: Qualitative comparison for the supervision-target ablation on GoPro-7. Best viewed zoomed in.

Figure 7: Additional blur-to-video comparisons on GoPro, trained on odd frames only.

Figure 8: Qualitative comparison for the loss-term ablation on GoPro-7 (sequence view). Each row shows a recovered $7$ -frame sequence under a different combination of training losses. Best viewed zoomed in.

Figure 9: Qualitative comparison for the loss-term ablation on GoPro (center-frame deblurring view). Each column shows the recovered center sharp frame under a different combination of training losses. Best viewed zoomed in.

Figure 10: Additional single-image deblurring comparisons. Two rows per dataset, GoPro, HIDE, and RealBlur-J, RealBlur-R (top to bottom).

Figure 11: Additional blur-to-video comparisons on GoPro-7.

Asset	Type	Citation	License
GoPro_Large_All	Dataset	[ 28 ]	Research use (per authors)
HIDE	Dataset	[ 35 ]	Research use (per authors)
RealBlur-J / RealBlur-R	Dataset	[ 32 ]	CC BY-NC-SA 4.0
FLUX.2-klein	Pretrained model	[ 3 ]	FLUX.2 Community License (non-commercial)
Stable Diffusion (referenced)	Pretrained model	[ 33 ]	CreativeML Open RAIL-M
RAFT	Optical flow	[ 38 ]	BSD 3-Clause

Appendix

Table 10: Existing assets used in this work. Datasets, pretrained models, and software with their licenses.

Explore similar work

RealVDeblur: One-Step Diffusion for Generalizable Real-World Video Deblurring
Jul 22, 2026Renbiao Jin, Mingxin Yang, Yutian Chen +8Gaussian Deblurring Video Restoration
1Shanghai Jiao Tong University · 2Shanghai AI Laboratory · 3CUHK MMLab +2
Restoration-Aligned Generative Flow Models for Blind Motion Deblurring
May 9, 2026Insoo Kim, Jinwoo ShinGaussian Deblurring Flow Models
1NAVER Cloud · 2KAIST AI
SloMoDeblur: A Large-Scale Smartphone Image Deblurring Dataset
Jun 24, 2025Syed Mumtahin Mahmud, Mahdi Mohd Hossain Noki, Prothito Shovon Majumder +4Gaussian Deblurring Smartphone Photographs
Cognitive Agents and Interaction Lab, Department of Computer Science and Engineering, University of Dhaka, Dhaka-1000, Bangladesh

Jul 22, 2026, cs.CVJ/K move · Enter open · S save

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

Renbiao Jin, Mingxin Yang, Yutian Chen, Junhao Zhuang+7

1Shanghai Jiao Tong University · 2Shanghai AI Laboratory · 3CUHK MMLab+2

Gaussian Deblurring Video Restoration

Open paper AskPDF arXiv

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

Abstract

Explore similar work

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

Restoration-Aligned Generative Flow Models for Blind Motion Deblurring

SloMoDeblur: A Large-Scale Smartphone Image Deblurring Dataset