While many diffusion models have achieved impressive results in real-world video super-resolution (Real-VSR) by generating rich and realistic details, their reliance on multi-step sampling leads to slow inference. One-step networks like SeedVR2, DOVE, and DLoRAL alleviate this through condensing generation into one single step, yet they remain heavy, with billions of parameters and multi-second latency. Recent adversarial diffusion compression (ADC) offers a promising path via pruning and distilling these models into a compact AdcSR network, but directly applying it to Real-VSR fails to balance spatial details and temporal consistency due to its lack of temporal awareness and the limitations of standard adversarial learning. To address these challenges, we propose an improved ADC method for Real-VSR. Our approach distills a large diffusion Transformer (DiT) teacher DOVE equipped with 3D spatio-temporal attentions, into a pruned 2D Stable Diffusion (SD)-based AdcSR backbone, augmented with lightweight 1D temporal convolutions, achieving significantly higher efficiency. In addition, we introduce a dual-head adversarial distillation scheme, in which discriminators in both pixel and feature domains explicitly disentangle the discrimination of details and consistency into two heads, enabling both objectives to be effectively optimized without sacrificing one for the other. Experiments demonstrate that the resulting compressed AdcVSR model reduces complexity by 95% in parameters and achieves an 8× acceleration over its DiT teacher DOVE, while maintaining competitive video quality and efficiency.
Figures & tables
Figure 1: Comparison of methods in compressing diffusion networks for Real-VSR. (a) Traditional ADC ( Chen et al., 2025a ) distills an SD network with 2D spatial attentions into a pruned student using a single adversarial signal without temporal modeling, suffering from frame flickering. (b) Our improved ADC distills a larger DiT-based teacher with heavier 3D spatio-temporal attention into the same 2D student, augmented by 1D temporal convolutions, using dual-head discriminators D in pixel and feature domains. Through disentangling the discriminations of detail richness and temporal consistency into different heads, it balances the optimization of both.
Figure 2: Illustration of the proposed improved ADC method, with application to compressing DOVE (teacher) into AdcVSR model (student). (a) We augment the 2D AdcSR Real-ISR network ( Chen et al., 2025a ) , consisting of pruned SD2.1 UNet and VAE decoder, through inserting 1D temporal residual blocks (RBs) after each 2D spatial RB and Transformer block (TB), enabling temporal modeling, while maintaining efficiency for Real-VSR. The resulting AdcVSR network is then fully trained end-to-end via adversarial distillation from DOVE. (b) For adversarial learning, we design dual-head discriminators with pretrained backbones for feature extractions. Each discriminator uses 2D and 1D convolutions, followed by two linear projection heads at the tail, to separately evaluate detail richness and temporal consistency. Training is guided by five curated types of video and image data (1-5) with head-specific labels, achieving a balanced optimization of details and consistency.
Method
RealBasicVSR
Upscale-A-Video
MGLD-VSR
STAR
SeedVR2
DOVE
DLoRAL
PiSA-SR
AdcSR
HYPIR
AdcVSR (Ours)
Test Dataset: UDM10 (Synthetic)
PSNR↑
24.39
23.03
24.28
22.61
25.92
26.00
22.49
23.21
23.39
22.55
25.36
SSIM↑
0.7376
0.6189
0.7491
0.6534
0.7674
0.7805
0.7130
0.6799
0.6772
0.6995
0.7697
LPIPS↓
0.3283
0.4218
0.3103
0.5055
0.2653
0.2645
0.3201
0.3658
0.3781
0.3736
0.3065
DISTS↓
0.2078
0.2360
0.1909
0.2665
0.1532
0.1732
0.2066
0.2213
0.2287
0.2125
0.2112
MANIQA↑
0.5725
0.5331
0.5558
0.3468
0.5232
0.5133
0.5679
0.6257
0.5696
0.5856
0.5793
Table 1: Quantitative comparison of Real-VSR performance. Inference time is measured on an NVIDIA H20 GPU for generating a 25-frame video at spatial resolution 512×512 . The best, second- best, and third-best results are labeled in bold red , underlined blue , and italic green , respectively.
Figure 3: Qualitative comparison of Real-VSR performance on 13th frames of two videos: “028” from RealVSR (top) and “016” from VideoLQ (bottom). Temporal profiles are provided below each frame, obtained by taking slices along the width-temporal plane at the vertical centers of the frames.
Figure 4: Performance comparison among diffusion-based Real-VSR methods in temporal consistency and complexity (parameter number and inference time) (see Tab. 1 ). AdcVSR attains the lowest warping error Ewarp∗ , the second-lowest parameter number, and the second-highest inference speed. Bubble colors represent method types: green for multi-step, blue for one-step, and red for AdcVSR.
Figure 6
Method
CLIPIQA↑
Ewarp∗ ↓
Single-Head, Dual-Domain
0.6745
6.32
Dual-Head, Single-Domain
0.6421
3.59
Dual-Head, Dual-Domain (Ours)
0.6861
2.22
Table 3: Comparison of discriminators on YouHQ40.
Method
PSNR↑
LPIPS↓
MUSIQ↑
No Adversarial Loss
23.97
0.3596
54.33
No Teacher (HR GT Only)
24.85
0.3641
50.32
SeedVR2 as Teacher
23.24
0.3489
60.74
DLoRAL as Teacher
23.08
0.3554
54.61
DOVE as Teacher (Ours)
23.81
0.3337
61.48
Table 4: Comparison of distillation setups on MVSR4x.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 9Figure 10
Training Configuration
LPIPS ↓
CLIPIQA ↑
Ewarp∗↓
No xvideo∗ and No ximage∗ (Types 3 & 5)
0.3430
0.6521
5.92
Real Videos as Detail-Real (Label “1”)
0.3152
0.6703
1.98
Teacher Videos Replacing Real Videos
0.3224
0.6652
2.47
Full Scheme (Ours)
0.3065
0.6818
1.67
Appendix
Table 7: Ablation of curated data for D on UDM10.
Method
RealBasicVSR
Upscale-A-Video
MGLD-VSR
STAR
SeedVR2
DOVE
DLoRAL
PiSA-SR
AdcSR
HYPIR
AdcVSR (Ours)
Test Dataset: SPMCS (Synthetic)
PSNR↑
20.69
19.83
20.83
20.32
20.89
20.94
20.35
20.12
20.14
18.13
20.83
SSIM↑
0.5056
0.4301
0.5201
0.4897
0.5234
0.5221
0.4955
0.4674
0.4702
0.4284
0.5055
LPIPS↓
0.3886
0.4336
0.3583
0.5657
0.3122
0.3165
0.3531
0.3662
0.3695
0.3969
0.3282
DISTS↓
0.2334
0.2466
0.2198
0.2957
0.1823
0.1896
0.2164
0.2286
0.2304
0.2251
0.2036
MANIQA↑
0.5977
0.5862
0.5941
0.3654
0.6289
0.5774
0.6066
0.6561
0.6332
0.6961
0.6500
Appendix
Table 8: Quantitative comparison of Real-VSR performance on 4 additional benchmark datasets.
Figure 8: Qualitative comparison of Real-VSR performance on three real videos: the 4th frame of “029” (top), the 2nd frame of “090” (middle), and the 15th frame of “247” (bottom) from RealVSR.
Real-time video super-resolution requires high spatio-temporal fidelity under strict latency constraints, challenging diffusion models due to their iterative sampling cost and limited temporal coordination. We propose a streaming-aware framework that adapts pretrained single-image latent diffusion models for efficient video super-resolution (VSR) by exploiting the sequential structure of video streams. Our Cross-Step Attention mechanism reuses intermediate denoising features across adjacent frames and diffusion steps, enabling temporal information exchange without explicit temporal modeling. We further introduce Trajectory-Coupled Diffusion Scheduling, which aligns adjacent diffusion states and provides cleaner intermediate representations for cross-step conditioning, improving temporal coherence. These components are integrated into a streaming inference pipeline that incrementally propagates latent states across frames, reducing the effective computational complexity from O(N⋅S) to O(N+S) for N frames and S diffusion steps. Experiments on REDS4 and YouHQ40-Test demonstrate improved perceptual quality and temporal realism while maintaining frame-wise stability. Our method achieves over 40 FPS at 512×512 resolution after cold start, enabling real-time VSR without explicit temporal modeling.
Harris Partaourides, Sotirios Chatzis
Ethical AI Novelties Limassol, Cyprus · Cyprus University of Technology Limassol, Cyprus
Real-world image super-resolution (Real-ISR) focuses on recovering high-quality images from low-resolution inputs that suffer from complex degradations like noise, blur, and compression. Recently, diffusion models (DMs) have shown great potential in this area by leveraging strong generative priors to restore fine details. However, their iterative denoising process incurs high computational overhead, posing challenges for real-time applications. Although one-step distillation methods, such as OSEDiff and TSD-SR, offer faster inference, they remain fundamentally constrained by their large, over-parameterized model architectures. In this work, we present TinySR, a compact yet effective diffusion model specifically designed for Real-ISR that achieves real-time performance while maintaining perceptual quality. We introduce a Dynamic Inter-block Activation and an Expansion-Corrosion Strategy to facilitate more effective decision-making in depth pruning. We achieve VAE compression through channel pruning, attention removal and lightweight SepConv. We eliminate time- and prompt-related modules and perform pre-caching techniques to further speed up the model. TinySR significantly reduces computational cost and model size, achieving up to 5.68x speedup and 83% parameter reduction compared to its teacher TSD-SR, while still providing high quality results.
Linwei Dong, Qingnan Fan, Yuhang Yu +4
Zhejiang University · Vivo Mobile Communication Co. Ltd · Zhejiang Lab
Video super-resolution (VSR) seeks to reconstruct high-resolution frames from low-resolution inputs. While diffusion-based methods have substantially improved perceptual quality, extending them to video remains challenging for two reasons: strong generative priors can introduce temporal instability, and multi-frame diffusion pipelines are often too expensive for practical deployment. To address both challenges simultaneously, we propose InstaVSR, a lightweight diffusion framework for efficient video super-resolution. InstaVSR combines three ingredients: (1) a pruned one-step diffusion backbone that removes several costly components from conventional diffusion-based VSR pipelines, (2) recurrent training with flow-guided temporal regularization to improve frame-to-frame stability, and (3) dual-space adversarial learning in latent and pixel spaces to preserve perceptual quality after backbone simplification. On an NVIDIA RTX 4090, InstaVSR processes a 30-frame video at 2K×2K resolution in under one minute with only 7 GB of memory usage, substantially reducing the computational cost compared to existing diffusion-based methods while maintaining favorable perceptual quality with significantly smoother temporal transitions.
Jintong Hu, Bin Chen, Zhenyu Hu +3
Insta360 Research · Peking University · Wuhan University