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

One blurry photo, every moment inside it

Pick a moment.

Drag along the exposure timeline. In moment mode the model returns the sharp frame $S(\tau)$ at the chosen instant; in interval mode it returns the interval-mean blur $B(s,t)$, i.e. the photo a camera would have taken had its shutter been open only from $s$ to $t$.

Blurry input B(0, 1)

Sharp frame S(τ)

shutter opens (0)normalized exposure timeshutter closes (1)

Teaser figure (paper Fig. 1): deblurring, blur-to-video and arbitrary-time recovery from one blurry image — Given one motion-blurred image, PickMoment answers three kinds of queries with the same network: **(a)** the sharp frame at the exposure center, **(b)** an $N$-frame sharp video, and **(c)** the sharp frame at any moment inside the exposure.

+1.23 dB

PSNR_p over Blur2Vid on GoPro-7 blur-to-video

+1.35 dB

mean PSNR over Blur2Vid at 33 timestamps per exposure (X4K1000FPS, unseen in training)

≈25×

faster than Blur2Vid: 7.90 s for a full 7-frame sequence (one forward pass per frame)

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-prior-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 with one forward pass per output frame and no iterative sampling.

Key idea

One function, many queries.

A motion-blurred photo is the average of the sharp scene over the exposure. PickMoment learns this averaging for every sub-interval, not only for the full exposure:

$$ B(s,t) \;=\; \frac{1}{t-s}\int_{s}^{t} S(u)\,\mathrm{d}u $$

for any $0 \le s < t \le 1$ in normalized exposure time. The observed blurry image is $B(0,1)$, and as the interval shrinks, $B(s,t)$ converges to the sharp frame $S(\tau)$ at that instant. Every task becomes a query to the same network:

Single-image deblurring

$S(\tfrac12)$

A zero-width query at the exposure center. One forward pass.

Blur-to-video

$\{S(\tau_i)\}_{i=0}^{N-1},\ \tau_i = \tfrac{i}{N-1}$

$N$ zero-width queries on the same input: one forward pass per output frame, issued as a single batch.

Pick-a-moment

$S(\tau)$ or $B(s,t)$

Any instant inside the exposure, or the blur of any sub-exposure. The other two tasks are special cases of this one.

Relation to MeanFlow

MeanFlow learns the time-average of an instantaneous velocity, so that one network evaluation replaces integration over an interval. PickMoment applies the same interval-averaging operator to a different quantity: sharp frames over exposure time. What carries over is the structure of an interval mean, namely additivity across sub-intervals and the zero-width limit. The stochastic noise-to-data sampling of MeanFlow does not: PickMoment builds on a pretrained generative prior but runs deterministically, so it does not retain the sampling-based generative capability. Because the interval mean of sharp frames is the observation itself, it can also be supervised directly.

Method

An image prior that answers a video question.

The network $f_\theta$ is a diffusion transformer built on the FLUX.2-klein backbone, with a frozen VAE (encoder $\mathcal{E}$, decoder $\mathcal{D}$). The blurry input is encoded as $z_\text{blur} = \mathcal{E}(x_\text{blur})$, and $f_\theta(z_\text{blur}, s, t)$ predicts the latent of $B(s,t)$ as a residual added to $z_\text{blur}$. The interval replaces FLUX's original time conditioning: the triple $(s, t, t-s)$ is sinusoidally embedded, projected by a shared MLP, and injected through the same pathway as the original timestep. An image model thus solves a video task, with no video pretraining. Training uses three supervisions derived from the blur integral:

Paper Fig. 2: conceptual illustration of the training strategy — reconstruction, additivity and sharp-frame losses — Conceptual illustration of the training strategy of PickMoment (paper Fig. 2).

Reconstruction $\mathcal{L}_\text{recon}$

Supervises $B(s,t)$ on sampled sub-intervals with the empirical mean of the available subframes.

Additivity $\mathcal{L}_\text{add}$

For every split point $m$, an interval mean is the length-weighted average of its two parts: \[ \begin{aligned} B(s,t) = {} & \tfrac{m-s}{t-s}\,B(s,m) \\ & + \tfrac{t-m}{t-s}\,B(m,t). \end{aligned} \] Enforcing it ties all sub-interval predictions on one input into a single consistent trajectory.

Sharp frame $\mathcal{L}_\text{sharp}$

Anchors zero-width queries $B(\tau,\tau)$ to the sharp frame $S(\tau)$, supervised in pixel space after decoding.

Results

Blur-to-Video Results

GoPro-7

Method	PSNR_p ↑	SSIM_p ↑	LPIPS_p ↓	FVD ↓	EPE ↓
Jin et al.	25.23	0.8190	0.084	235.53	3.38
MotionETR	26.54	0.8825	0.015	94.90	1.46
Blur2Vid	30.01	0.9359	0.012	21.46	0.39
PickMoment (Ours)	31.24	0.9388	0.0086	27.40	0.86

Bidirectional patch-based metrics following the Blur2Vid protocol. Baseline numbers are from the respective papers.

Trade-off

PickMoment leads on per-frame fidelity (PSNR_p, SSIM_p, LPIPS_p). Blur2Vid leads on FVD and EPE, the two metrics that judge the sequence as a whole. This follows from the prior: Blur2Vid fine-tunes a video diffusion model (CogVideoX-2B) and samples the whole sequence jointly, while PickMoment uses an image-only prior with no cross-frame attention and answers each query independently. Against every other baseline, PickMoment's FVD and EPE are far ahead.

Temporal consistency

Sequence	warp L1	warp PSNR	vs. ground truth
Ground truth (natural motion)	0.0096	35.23	—
PickMoment (Ours)	0.0089	36.49	closest to GT
Blur2Vid	0.0072	37.77	smoother than GT
Jin et al.	0.0106	34.52	rougher than GT

Each recovered frame is warped onto the next along RAFT flow and the photometric residual is measured, on all 1,744 GoPro-7 clips. PickMoment's frame-to-frame change matches natural motion, with no measurable flicker.

Denser timestamps on unseen data

$t/T$	0.000	0.125	0.250	0.375	0.500	0.625	0.750	0.875	1.000	33-pt mean
Blur2Vid	21.31	20.30	19.99	22.92	21.46	22.47	21.95	21.53	21.30	21.60
PickMoment (Ours)	22.09	22.39	23.29	23.65	24.07	23.61	22.89	21.82	21.60	22.95

PSNR on X4K1000FPS (X-TEST, 15 clips). Each blur averages 33 consecutive 1000-fps frames, leaving ground truth at 33 timestamps per exposure. Both models are applied zero-shot from their GoPro-trained weights; PickMoment leads at every one of the 33 timestamps (9 shown).

Results · Single-image deblurring

The same model, queried at τ = ½.

Drag the slider to compare the blurry input with PickMoment's sharp frame at the exposure center (GoPro test set, full resolution).

Comparison with FideDiff

Method	GoPro				HIDE
Method	PSNR ↑	SSIM ↑	LPIPS ↓	DISTS ↓	PSNR ↑	SSIM ↑	LPIPS ↓	DISTS ↓
FideDiff	28.79	0.9148	0.0831	0.0525	27.28	0.8775	0.1068	0.0647
PickMoment (Ours)	29.43	0.9253	0.0802	0.0429	27.52	0.8889	0.1043	0.0551

Comparison with FideDiff, the strongest generative-prior-based baseline. Across GoPro, HIDE, RealBlur-J and RealBlur-R, PickMoment is ahead on 12 of 16 metric cells; the full table with regression-based methods is in the paper. Deblurring uses the same model as blur-to-video, queried at $\tau = \tfrac12$.

Analysis

What does the interval-mean formulation add?

Same backbone, different target. With backbone, data and budget fixed, and only a latent reconstruction loss (no sharp-frame or perceptual loss), predicting the interval mean beats predicting a single sharp frame by 0.73 dB and a fixed set of $N$ frames by 0.69 dB in deblurring PSNR (0.62 / 0.52 dB in PSNR_p).

Sharp-frame loss alone. Training with $\mathcal{L}_\text{sharp}$ only, so that no interval-mean supervision remains, nearly matches the full objective on deblurring but falls well behind on the video metrics. The formulation's contribution is temporal.

GoPro	deblur PSNR	SSIM	LPIPS	DISTS	PSNR_p	SSIM_p	LPIPS_p	FVD	EPE
$\mathcal{L}_\text{sharp}$ only	29.06	0.9208	0.0812	0.0438	30.17	0.9186	0.0104	38.07	1.33
Full objective	29.43	0.9253	0.0802	0.0429	31.24	0.9388	0.0086	27.40	0.86

Citation

BibTeX

@inproceedings{shin2026pickmoment,
    title={PickMoment: Continuous-Time Single-Image-to-Video via Learning Deblurring and Blur-to-Video},
    author={Junseong Shin and Hyeonsu Jo and Daehyun Kim and Tae Hyun Kim},
    booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
    year={2026}
}