Diffusion-based video super-resolution (VSR) has recently achieved remarkable fidelity but still suffers from prohibitive sampling costs. While distribution matching distillation (DMD) can accelerate diffusion models toward one-step generation, directly applying it to VSR often results in training instability alongside degraded and insufficient supervision. To address these issues, we propose DUO-VSR, a three-stage framework built upon a Dual-Stream Distillation strategy that unifies distribution matching and adversarial supervision for one-step VSR. Firstly, a Progressive Guided Distillation Initialization is employed to stabilize subsequent training through trajectory-preserving distillation. Next, the Dual-Stream Distillation jointly optimizes the DMD and Real-Fake Score Feature GAN (RFS-GAN) streams, with the latter providing complementary adversarial supervision leveraging discriminative features from both real and fake score models. Finally, a Preference-Guided Refinement stage further aligns the student with perceptual quality preferences. Extensive experiments demonstrate that DUO-VSR achieves superior visual quality and efficiency over previous one-step VSR approaches.
Figures & tables
Figure 2: (a) Effect of initialization on the stability of the second-stage training. The proposed progressive guided distillation initialization leads to more stable loss and gradient norm trends during the second-stage distillation. (b) Compared with the fake score model, the real score model occasionally produces outputs that are spatially shifted relative to the inputs (highlighted in green boxes in the first two cases) or contain artifacts ( blue boxes in the third case), leading to degraded supervision propagated to the student model.
Figure 3: Overview of our three-stage distillation framework. (a) We initialize the student model with trajectory-preserving Progressive Guided Distillation, which consists of CFG Distillation and Progressive Distillation steps. (b) The core of our method, Dual-Stream Distillation, jointly optimizes the DMD and RFS-GAN streams through alternating Student Update and Auxiliary Update, providing reliable and sufficient supervision. (c) In the final stage, we construct a generated preference dataset and apply DPO-based Preference-Guided Refinement to enhance perceptual quality.
Figure 4: Visual comparison on synthetic (YouHQ40), real-world (VideoLQ) and AIGC (AIGC60) datasets. Zoom in for details.
Datasets
Metrics
RealViformer
VEnhancer
MGLD
UAV
STAR
DLoRAL
DOVE
SeedVR2-7B
DUO-VSR
PSNR ↑
21.34
19.92
21.89
19.67
21.56
22.87
22.69
23.08
22.90
SSIM ↑
0.601
0.523
0.642
0.538
0.613
0.674
0.694
0.685
0.691
LPIPS ↓
0.394
0.417
0.348
0.424
0.386
0.331
0.318
0.302
0.315
NIQE ↓
4.67
4.36
3.79
3.87
6.17
3.96
4.79
4.68
3.59
MUSIQ ↑
62.13
61.10
63.43
58.52
29.08
61.42
65.27
64.74
66.91
CLIP-IQA ↑
0.3443
0.3398
0.4333
0.4582
0.3827
0.5218
0.4922
0.5073
0.5459
Table 1: Quantitative comparisons on benchmarks, including synthetic (SPMCS [ 55 ] , UDM10 [ 81 ] , YouHQ40 [ 96 ] ), real-world (VideoLQ [ 4 ] ), and AIGC (AIGC60) videos. The best and second performances are marked in red and blue respectively.
Figure 5: Comparison of temporal consistency. Extracted and stacked along the blue line in the width–temporal plane.
Metric
UAV
MGLD
VEnh.
STAR
DOVE
DLoRAL
SeedVR2
DUO-VSR
Step
30
50
15
15
1
1
1
1
Time (s)
382.1
956.7
404.5
200.4
66.7
76.6
89.7
11.3
Params (B)
0.7
1.4
2.0
2.0
5.6
0.9
8.2
1.3
Table 2: Inference efficiency comparison. Measured on a single GPU using a 21-frame 1920×1080 video. The model parameters are counted only for the generator part.
Figure 6: Visual comparison of ablation on three stage distillation. Experiment indices refer to Tab. 3 . Zoom in for details.
Exp.
Variants
Metrics
I
II
III
NIQE
MUSIQ
CLIPIQA
DOVER
Base
4.31
63.46
0.4712
87.98
(a)
✓
5.45
58.97
0.408
86.49
(b)
✓
✓
4.64
63.36
0.487
88.01
(c)
✓
✓
5.11
60.22
0.423
87.63
(d)
✓
✓
✓
4.42
63.68
0.489
88.15
Table 3: Ablation on Three Stage Distillation.
Setting
NIQE
MUSIQ
CLIPIQA
DOVER
Component-level
DMD only
4.99
61.46
0.432
87.38
RFS-GAN only
5.32
62.64
0.427
87.53
Dual-Stream (Joint)
4.42
63.68
0.489
88.15
Optimization-level
Sequential DMD → GAN (Seq.)
5.17
62.76
0.419
87.67
Table 4: Ablation on Dual-Stream Distillation Strategy. “Joint” and “Seq.” denote different optimization schemes.
Figure 7: Visual comparison of ablation on Dual-Stream Distillation Strategy. The orange and red boxes show spatial comparison in the LR. The blue box shows the temporal profile along the blue line in the LR. Zoom in for details.
Figure 8: Examples of preferred and less-preferred samples in the constructed preference dataset. Zoom in for details.
Figure 9: Noise-perturbed samples stabilize adversarial training and suppress artifacts. Zoom in for details.
Figure 10: Discriminator features from the real and fake score models used for the RFS-GAN loss computation, reduced to three dimensions via t-SNE [ 39 ] for visualization.
Method
NIQE ↓
MUSIQ ↑
CLIPIQA ↑
DOVER ↑
Real Score Model only
4.71
62.79
0.456
87.95
Fake Score Model only
4.98
62.98
0.475
87.76
RFS-GAN
4.64
63.36
0.487
88.01
Table 5: Ablation study on the discriminator design of RFS-GAN.
Figure 11: Visual comparison with base VSR model. Zoom in for details.
Metric
UltraVSR
FlashVSR
DUO-VSR
NIQE ↓
5.58
4.67
4.42
MUSIQ ↑
58.23
63.11
63.68
CLIP-IQA ↑
0.4434
0.4690
0.4886
DOVER ↑
86.45
87.49
88.15
Ewarp∗↓
1.54
1.76
1.08
Time (s)
126.5
10.7
11.3
Table 6: Quantitative comparison on the AIGC60 dataset.
Figure 12: Visual comparison of different VSR methods. DUO-VSR consistently reconstructs finer textures. Zoom in for details.
Method
Overall Quality
Visual Fidelity
Visual Quality
DUO-VSR
0%
0%
0%
Our Base VSR model
-1.3%
-3.7%
+2.3%
SeedVR2-7B
-32.7%
-13.3%
-39.2%
DOVE
-29.3%
-8.0%
-36.2%
DLoRAL
-34.0%
-10.8%
-37.8%
UltraVSR
-39.8%
-16.7%
-43.3%
Table 7: Blind user study results based on GSB test.