PaperlayerView exact revision on arXiv

Source-linked AI summary

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

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

arXiv:2610.01279v1cs.CVcs.AI

TL;DR

Existing learning-based methods typically predict a single exposure-center frame or a fixed set of sharp frames instead of modeling the blur integral. PickMoment learns interval-mean blur over arbitrary exposure sub-intervals with one deterministic model, achieving strong reported deblurring and blur-to-video results in a single forward pass.

  • Problem

    Existing deblurring methods typically recover only an exposure-center frame, while blur-to-video methods predict sharp frames at fixed discrete times rather than modeling the temporal structure of blur.

    Source: §1 · p. 2

  • Method

    PickMoment trains one deterministic model to predict interval-mean blur over arbitrary exposure sub-intervals, with empirical reconstruction, additivity, and zero-interval sharp-frame supervision.

    Source: §5 · p. 9

  • Results

    PickMoment achieves generative-based deblurring state-of-the-art on GoPro and HIDE, competitive RealBlur performance, and the highest per-frame fidelity on GoPro-7 blur-to-video in one forward pass.

    Source: §5 · p. 9

  • Takeaways & Limitations

    A single trained model supports deblurring, blur-to-video, and continuous-time pick-a-moment recovery as different queries.

    Source: §5 · p. 9

    The additivity constraint is applied to latent outputs using the VAE encoder’s approximate linearity, rather than exact pixel-space additivity.

    Source: §4.1 · p. 6

Abstract

from arXiv · show

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.

1 Introduction

Motion blur integrates sharp frames over an exposure, but existing deblurring and blur-to-video methods predict sharp content at one or fixed discrete times. PickMoment instead predicts interval-mean blur over arbitrary sub-intervals, using integral-derived constraints to unify several temporal queries in one model.

1 Introduction

  • The framework treats the blurred image as temporal integration of sharp frames over a finite exposure, a physically grounded interpretation also used for blur synthesis and deblurring.The formulation follows Nah et al. [28].Source: §1 · p. 1
  • Existing deblurring methods typically recover only the exposure-center frame, while blur-to-video methods predict fixed sharp-frame sequences rather than directly modeling the blur integral.This leaves temporal structure encoded in blur underused.Source: §1 · p. 2
  • PickMoment predicts blur observations over arbitrary exposure sub-intervals, using the temporal integral itself as the learning target rather than directly predicting sharp frames.Its MeanFlow-inspired strategy uses integral-derived identities for self-consistency across sub-intervals.Source: §1 · p. 2
  • A single PickMoment model supports center-frame deblurring, blur-to-video generation, and arbitrary-moment recovery as different queries.The arbitrary-time sharp frame is obtained as the zero-interval limit.Source: §1 · p. 2
  • PickMoment achieves state-of-the-art performance among generative-based deblurring methods on GoPro and HIDE while remaining competitive with restoration-based baselines.It also enables one-step continuous-time blur-to-video generation with reduced computational cost compared with video diffusion approaches.Source: §1 · p. 3

2 Related Work

Prior work addresses single-frame deblurring, blur-to-video reconstruction, and continuous-time average-velocity modeling through distinct approaches. PickMoment draws on the latter formulation to predict interval averages of sharp frames rather than committing to isolated frames.

Single-image motion deblurring

Blur-to-video

  • Blur-to-video methods address directional ambiguity through order-invariant losses [15], motion priors [31] [2], motion estimation [48], direction guidance [49], latent ordering [30], or video-diffusion sampling [37].Despite these varied strategies, they predict sharp content directly rather than modeling the integral structure of motion blur.Source: §2 · p. 3

Average-velocity Flow Matching

  • MeanFlow predicts average velocity over an interval to encode a trajectory in one arbitrary-interval prediction, and SplitMeanFlow [11] adds an algebraic additivity objective.PickMoment adopts this perspective by predicting the temporal average of sharp frames over an exposure sub-interval.Source: §2 · p. 3

3 Method

PickMoment represents blur as interval-mean sharp content, using integral-derived constraints to train one model for arbitrary temporal queries. The model is supervised by empirical subframe means, interval additivity, and the zero-interval sharp-frame limit, then queried for deblurring, video, or a selected moment.

3.1 Preliminary: MeanFlow

  • MeanFlow predicts average velocity over intervals; SplitMeanFlow uses a split-interval identity as a JVP-free training target, reducing per-step compute and memory.The split target is formed by combining predictions over two adjacent intervals.Source: §3.1 · pp. 3–4

3.2 Continuous-Time Blur Reformulation

  • PickMoment defines B(s, t) as average sharp content over any exposure sub-interval; adjacent interval blurs combine by convex weighting, and a zero-width interval yields the sharp frame.This makes arbitrary-time deblurring and continuous-time video or pick-a-moment recovery queries of the same representation.Source: §3.2 · p. 4

3.3 Training Strategy

  • Training combines empirical subframe-mean reconstruction, additivity across overlapping intervals, and direct sharp-frame supervision at the zero-interval boundary.The additivity constraint samples an intermediate timestamp to propagate point-wise supervision across the interval domain; the sharp-frame loss uses latent and perceptual distances.Source: §3.3 · pp. 5–6

3.4 Inference

  • One trained model supports deblurring, blur-to-video, and user-selected-moment recovery by changing the interval query, with deterministic outputs in one forward pass and no retraining or iterative sampling.Video queries use uniformly spaced timestamps; pick-a-moment queries accept any timestamp in the exposure.Source: §3.4 · p. 6

4 Experiments

Experiments evaluate one GoPro-trained PickMoment model on single-image deblurring and blur-to-video, then test interval prediction, unseen-timestamp generalization, and loss contributions. The reported results show strong generative-method deblurring performance and highest per-frame fidelity on GoPro-7.

4.1 Experimental Setup

  • Training uses GoPro’s ordered 7–13-frame exposures and evaluates one unfine-tuned model on four deblurring datasets and GoPro-7.Single-image metrics use the center sharp frame; blur-to-video also reports patch-based metrics, FVD, and optical-flow EPE. Latent additivity relies on approximate VAE-encoder linearity.Source: §4.1 · p. 6

4.2 Single-Image Deblurring Results

  • PickMoment is state-of-the-art among generative methods on GoPro and HIDE across four metrics, exceeding FideDiff by +0.64 dB PSNR on GoPro and +0.24 dB on HIDE.RealBlur-J/R results are competitive with FideDiff; PickMoment is better on 12 of 16 evaluation metrics overall.Source: Table 1 · p. 7, §4.2 · p. 7

4.3 Blur-to-Video Results

  • PickMoment achieves the highest per-frame fidelity on GoPro-7, beating Blur2Vid by +1.23 dB in PSNRp and reducing LPIPSp by a relative 28%.Blur2Vid retains an advantage on FVD and EPE; PickMoment uses a single forward pass and yields roughly a 25× inference speedup.Source: Table 2 · p. 8, §4.3 · p. 8

4.4 Ablation Studies

  • Interval-mean supervision outperforms the two sharp-frame targets on every metric except FVD, where Point-N is marginally better.The comparison tests alternative supervision targets on GoPro.Source: Table 4 · p. 9, §4.4 · p. 9
  • Lsharp nearly halves LPIPS from the Lrecon-only value of 0.1455, while combining Ladd with Lsharp reduces FVD from 67.10 to 27.40 and EPE from 0.99 to 0.86.The authors describe Lsharp as enforcing per-frame sharpness and Ladd as propagating temporal coherence; Ladd is especially effective for blur-to-video.Source: Table 5 · p. 9, Figure 5 · p. 9, §4.4 · p. 9
  • When trained on odd-indexed subframes, PickMoment exceeds Point-1 at every unseen even timestamp, with +0.92 dB Mean PSNR and +0.023 Mean SSIM.The authors interpret this as smoother generalization across continuous time than direct sharp-frame prediction.Source: §4.4 · p. 9

5 Conclusion

PickMoment predicts interval-mean blur over arbitrary exposure sub-intervals, and one deterministic model serves deblurring, blur-to-video, and pick-a-moment queries. It performs strongly across the reported benchmarks in a single forward pass, while temporal smoothness remains a gap relative to iterative video-diffusion sampling.

5 Conclusion

  • PickMoment predicts interval-mean blur over arbitrary exposure sub-intervals, unifying single-image deblurring, blur-to-video, and pick-a-moment recovery in one deterministic model.The three supervision terms are empirical reconstruction, additivity across sub-intervals, and a sharp-frame loss at the zero-interval limit.Source: §5 · p. 9
  • PickMoment achieves state-of-the-art generative-based deblurring on GoPro and HIDE, competitive performance on RealBlur, and the highest per-frame fidelity on GoPro-7 blur-to-video.All results use a single forward pass.Source: §5 · p. 9
  • Temporal smoothness remains the main gap: iterative video-diffusion sampling provides stronger implicit regularization, and combining it with deterministic interval-mean prediction is a proposed next step.The paper notes that further limitations are discussed in Appendix G.Source: §5 · p. 9

A Training Details

The model uses a fine-tuned FLUX.2-klein transformer with interval conditioning and is trained on GoPro blurry images paired with ordered subframes. Training combines three losses, progressive learning, and a fixed optimization schedule.

Architecture and trainable parameters

  • The FLUX.2-klein-4B transformer is fully fine-tuned while its pretrained VAE remains frozen, and interval conditioning embeds (s, t, t−s) through the AdaLN pathway.The prediction is a residual displacement from the encoded blur input.Source: A Training Details · p. 12

Data pipeline

  • Training uses 2,103 GoPro blurry/sharp pairs, each with 7–13 consecutive subframes, with shared-coordinate random crops and horizontal flips.No color jitter or other augmentations are used.Source: A Training Details · p. 12

Loss sampling

  • The reconstruction loss samples ordered subframe pairs, with 25% of samples constrained to zero-length intervals; additivity samples an intermediate timestamp, while the sharp-frame loss samples subframe timestamps.The zero-interval branch and sharp-frame loss supervise the same sharp-frame target; EA-LPIPS follows the FideDiff implementation.Source: A Training Details · p. 12

Optimization and progressive learning

  • Training runs for 100K Adam iterations, with a learning rate warmed to 10^-4 then cosine-annealed to 10^-5 and progressive patch/batch settings of (256^2, 8), (384^2, 4), and (512^2, 2).Auxiliary loss weights warm up during the first 30K iterations; no EMA is used.Source: A Training Details · p. 12

Hardware and wall-clock time

B Empirical Validation of Latent-Space Additivity

The latent-space additivity assumption is tested on GoPro intervals and incurs an additional error comparable to the VAE round-trip floor. The section also reports a separate GoPro-7 center-frame evaluation protocol, which is not directly comparable to the standard GoPro test protocol.

B Empirical Validation of Latent-Space Additivity

  • The validation samples 2,983 triples from 11 GoPro test sequences, excludes very small intervals, and treats the zero-interval limit as separately supervised by sharp-frame and reconstruction losses.The reported measurements compare VAE round-trip error, pure additivity error, and combined error against the pixel target.Source: B Empirical Validation of Latent-Space… · p. 13
  • 42.70 dB mean PSNR for VAE round-trip reconstruction, 39.82 dB for the additive prediction against the round-trip output, and 37.40 dB against the true pixel target support applying additivity loss in latent space.Across 2,983 triples, the additivity error was essentially uncorrelated with interval width or split position.Source: B Empirical Validation of Latent-Space… · p. 13, Table 7 · p. 13

C Center Frame Quality on GoPro-7

  • Center-frame quality is evaluated on 1,744 GoPro-7 clips at 640 × 360 using frame 3 of each recovered seven-frame sequence, with full-image PSNR, SSIM, LPIPS, and DISTS.This protocol differs from the standard GoPro test split in averaging length, resolution, and test composition.Source: C Center Frame Quality on GoPro-7 · p. 13

D Additional Ablation Studies

Additional ablations show that the full loss objective has the best PSNR/SSIM across four datasets, while interval-mean supervision improves qualitative sharpness and generalization to held-out timestamps. The additivity and sharp-frame losses contribute distinct benefits to sequence coherence and frame sharpness.

Cross-dataset loss ablation

  • Across GoPro, HIDE, RealBlur-J, and RealBlur-R, the full objective achieves the best PSNR/SSIM; Ladd alone gives marginal gains over Lrecon, while Lsharp provides most LPIPS/DISTS improvement.The reported ordering is consistent across synthetic and real blur datasets.Source: D Additional Ablation Studies · p. 14, Table 9 · p. 14

Advantage of interval prediction

  • Interval-mean supervision yields sharper content and motion better aligned with ground-truth subframes than the B → S and B → {S(τi)} baselines in the GoPro-7 comparison.Figure 6 provides qualitative comparisons for this supervision-target ablation.Source: D Additional Ablation Studies · p. 14, Figure 6 · p. 15

Generalization to unseen timestamps

  • When trained on odd-indexed subframes, PickMoment recovers visibly sharper content at held-out even-indexed timestamps than the B → S baseline.The additional Figure 7 comparisons are trained on odd frames only.Source: D Additional Ablation Studies · p. 14, Figure 7 · p. 15

Effect of each loss term

  • Ladd promotes direction-consistent motion but leaves the center frame blurry, whereas Lsharp sharpens the center frame but leaves inter-frame jitter; the full objective combines both properties.The two identity-derived losses address orthogonal failure modes.Source: D Additional Ablation Studies · pp. 14–15

E Inference Cost

On a single NVIDIA RTX A6000 at 1280×640, PickMoment generates a 7-frame sequence in about 8 seconds, roughly 25× faster than Blur2Vid at its default 50 DDIM steps.

E Inference Cost

  • Roughly 25× faster than Blur2Vid, PickMoment generates a 7-frame video in one forward pass versus Blur2Vid’s default 50 DDIM steps.PickMoment averages 7.90 seconds over 10 runs after 3 warmup runs; Blur2Vid requires about 3.5 minutes per video.Source: E Inference Cost · p. 15
  • 7.90 seconds is PickMoment’s mean time to generate a 7-frame sequence across 10 runs after 3 warmup runs.Timing was measured on one NVIDIA RTX A6000 GPU at 1280×640 resolution.Source: E Inference Cost · p. 15
  • About 3.5 minutes per video is Blur2Vid’s inference time at the default 50 DDIM steps.The comparison uses the same reported 1280×640 timing setup.Source: E Inference Cost · p. 15

F Additional Visual Results

Additional qualitative comparisons show sharper details in PickMoment’s single-image deblurring results and sharper, direction-consistent motion in its GoPro-7 sequences.

F Additional Visual Results

  • Zoomed-in crops highlight texture preservation and motion-trajectory coherence across frames where applicable.The additional qualitative comparisons complement the main paper’s figures.Source: F Additional Visual Results · p. 15
  • Across GoPro, HIDE, and RealBlur, PickMoment recovers sharper edges and finer textures than the strongest Table 1 baseline, including on real blur.Figure 10 provides the additional single-image comparisons.Source: F Additional Visual Results · p. 15
  • On GoPro-7, PickMoment produces sharper per-frame content and a direction-consistent motion trajectory than Blur2Vid in the additional sequence comparisons.Figure 11 shows the blurry input, recovered sequences, and ground-truth subframes.Source: F Additional Visual Results · p. 15

G Limitations

PickMoment’s limitations concern training-data breadth, a trade-off between per-frame fidelity and cross-frame smoothness, and the assumed blur-formation model; the appendix also lists asset licenses.

G Limitations

  • Training only on GoPro_Large_All, PickMoment generalizes to HIDE and RealBlur but does not consistently outperform FideDiff on RealBlur-J.The authors identify broader training distributions as a potential source of improvement.Source: G Limitations · p. 15
  • PickMoment’s single-pass, independent interval predictions maximize per-frame quality, while Blur2Vid retains a small advantage on cross-frame smoothness metrics FVD and EPE.The comparison reflects Blur2Vid’s implicit temporal regularization from iterative sampling.Source: G Limitations · pp. 15–16
  • The assumed uniform temporal integral over a known exposure window fits GoPro synthesis but does not directly cover non-uniform exposures, severe rotational or radial blur, or camera-pipeline effects.The authors leave generalization to these regimes for future work.Source: G Limitations · p. 16

H Asset Licenses

  • Table 10 lists datasets, pretrained models, and software with their licenses; the authors state that assets are used according to their licenses and cited in the main paper.The asset table appears in the Asset Licenses section.Source: H Asset Licenses · p. 16, Table 10 · p. 18
Loading 2610.01279v1…