Existing continuous space-time video super-resolution methods often infer motion from only two endpoint frames and model the two motion directions independently, making them vulnerable to ambiguous correspondence and inconsistent trajectories. We introduce a continuous STVSR framework that combines long-context encoding, continuous HR feature lifting, and an endpoint-anchored PixelSpline. Long-range observations provide more reliable endpoint representations, while continuous upsampling constructs detailed appearance features directly on the requested HR grid. PixelSpline explicitly models each pixel's motion as a single continuous trajectory shared by both endpoints. Using this trajectory, our method warps endpoint colors into a coarse intermediate frame and aligns the HR features for residual reconstruction. Experiments across spatial and temporal upsampling settings demonstrate that the proposed framework offers a favorable quality--efficiency trade-off.
Figures & tables
Figure 1: Quality–latency trade-off on Adobe240 Average under S×4 spatial and T×8 temporal upsampling. Our method is 2.24× faster than V 3 , with a 0.90 dB PSNR difference. Latency is measured end-to-end on a single NVIDIA H100 GPU after warm-up.
Figure 2: Overview of our continuous STVSR framework. A bidirectional long-context encoder aggregates information from multiple LR frames to obtain context-enhanced endpoint representations (Sec.3.2) . Continuous feature upsampling maps the endpoint features onto the requested HR grid (Sec.3.3) , while an endpoint-anchored PixelSpline constructs a shared trajectory for RGB and feature transport at the queried time (Sec. 3.4) . Finally, a lightweight decoder predicts an HR residual that is added to the coarse warped frame to produce the reconstruction.
Method
GoPro Center
GoPro Average
Adobe Center
Adobe Average
REDS Center
REDS Average
Zooming SlowMo [ 27 ]
30.69 / 0.885
–
30.26 / 0.882
–
–
–
TMNet [ 28 ]
24.94 / 0.840
24.76 / 0.824
27.25 / 0.843
26.65 / 0.827
22.64 / 0.667
19.84 / 0.573
VideoINR [ 6 ]
30.26 / 0.879
29.41 / 0.867
29.92 / 0.875
29.27 / 0.865
23.25 / 0.650
19.85 / 0.558
MoTIF [ 4 ]
31.12 / 0.889
30.14 / 0.879
30.63 / 0.884
29.82 / 0.875
23.62 / 0.654
20.01 / 0.556
BF-STVSR [ 14 ]
31.22 / 0.890
30.26 / 0.881
30.83 / 0.888
30.12 / 0.880
23.55 / 0.646
19.95 / 0.549
V 3 [ 3 ]
32.98 / 0.923
32.27 / 0.919
32.91 / 0.922
32.29 / 0.917
24.00 / 0.671
20.36 / 0.573
Table 1: Quantitative comparison of C-STVSR reconstruction quality under S×4 spatial and T×8 temporal upsampling, reported as PSNR ↑ / SSIM ↑ . Best and second-best results are highlighted in bold and underlined, respectively.
Method
Latency (ms/frame) ↓
FPS ↑
Zooming SlowMo [ 27 ]
354.72
2.82
TMNet [ 28 ]
81.44
12.28
VideoINR [ 6 ]
360.58
2.77
MoTIF [ 4 ]
494.65
2.02
BF-STVSR [ 14 ]
417.51
2.40
V 3 [ 3 ]
912.21
1.10
Table 2: End-to-end inference efficiency on Adobe240 after one full-resolution warm-up pass. Latency includes input loading and preprocessing, device transfers, all method-specific inference operations, image conversion, PNG encoding, and disk writing.
Figure 3: Qualitative comparison of C-STVSR methods under S×4 spatial and T×8 temporal upsampling. The first three rows show results on GoPro, and the last two rows show results on REDS. Our method preserves fine details such as characters, texture patterns, and clothing structures while producing sharp and smooth boundaries. Best viewed zoomed in.
Method
TMNet
VideoINR
MoTIF
BF-STVSR
V 3
Ours
tOF ↓
9.944
9.495
9.391
9.237
8.923
8.171
Table 3: Temporal consistency on REDS under S×4 spatial and T×8 temporal upsampling, measured by tOF ↓ .
Figure 4: Qualitative comparison of temporal consistency. The red dashed line marks the vertical pixel column sampled from each frame to construct the temporal profile, while the red arrows indicate regions with noticeable deviations from the ground truth. Our method faithfully reconstructs fine details along nonlinear motion trajectories while suppressing motion-induced artifacts.
Setting
VideoINR [ 6 ]
MoTIF [ 4 ]
BF-STVSR [ 14 ]
V 3 [ 3 ]
Ours
S×4,T×1
25.97 / 0.775
26.21 / 0.780
26.28 / 0.782
27.15 / 0.819
26.78 / 0.808
S×1,T×8
19.76 / 0.462
28.86 / 0.864
29.42 / 0.810
31.15 / 0.862
29.01 / 0.890
Table 4: Decoupled spatial-only and temporal-only super-resolution results on Vid4 (PSNR ↑ / SSIM ↑ ). Temporal-only results are evaluated on center frames. Best and second-best results are highlighted in bold and underlined, respectively.