One-Step Diffusion Models have demonstrated promising capability and fast inference in video super-resolution (VSR) for real-world. Nevertheless, the substantial model size and high computational cost of Diffusion Transformers (DiTs) limit downstream applications. While low-bit quantization is a common approach for model compression, the effectiveness of quantized models is challenged by the high dynamic range of input latent and diverse layer behaviors. To deal with these challenges, we introduce LSGQuant, a layer-sensitivity guided quantizing approach for one-step diffusion-based real-world VSR. Our method incorporates a Dynamic Range Adaptive Quantizer (DRAQ) to fit video token activations. Furthermore, we estimate layer sensitivity and implement a Variance-Oriented Layer Training Strategy (VOLTS) by analyzing layer-wise statistics in calibration. We also introduce Quantization-Aware Optimization (QAO) to jointly refine the quantized branch and a retained high-precision branch. Extensive experiments demonstrate that our method has nearly performance to origin model with full-precision and significantly exceeds existing quantization techniques. Code is available at: https://github.com/zhengchen1999/LSGQuant.
Figures & tables
Figure 1: Visual comparison among the full-precision bfloat16 model, recent leading quantization method ViDiT-Q ( Zhao et al., 2025 ) , SVDQuant ( Li et al., 2025a ) , and our low-bit LSGQuant.
Figure 2: Performance comparisons on the synthetic MVSR4x dataset ( Wang et al., 2023b ) between different quantization methods in 4-bit and 6-bit scenarios. LSGQuant achieves leading scores on all evaluation metrics.
Figure 3: Overview of our LSGQuant. Firstly , we collect per-channel layer statistics by a single inference process. Next , we calculate layer sensitivity and estimate its importance to the final video output. Finally , we allocate training iterations by estimated sensitivities.
Figure 4: Input data sampled from the cross-attention query layer. (a) Statistical distribution of activation values. While the majority of values are concentrated near zero, a significant long-tail distribution exists. These extreme values would dominate the quantization range. (b) Distribution of input latent values across different blocks. The statistical characteristics of outliers vary with depth.
Dataset
Bit
Method
PSNR ↑
SSIM ↑
LPIPS ↓
DISTS ↓
CLIP-IQA ↑
MUSIQ ↑
MANIQA ↑
DOVER ↑
Ewarp∗↓
UDM10
W16A16
WAN ( Wan et al., 2025 )
23.74
0.7144
0.2886
0.2080
0.4973
63.85
0.3482
0.5494
0.93
W6A6
MinMax ( Jacob et al., 2018 )
23.27
0.7090
0.2987
0.2154
0.5205
63.73
0.3412
0.5135
0.95
SmoothQuant ( Xiao et al., 2023 )
23.02
0.7014
0.3011
0.2180
0.5237
64.39
0.3520
0.5579
0.89
QuaRot ( Ashkboos et al., 2024 )
23.66
0.7128
0.2912
0.2120
0.4990
63.78
0.3491
0.5403
0.92
ViDiT-Q ( Zhao et al., 2025 )
23.08
0.7024
0.3005
0.2167
0.5224
64.38
0.3499
0.5513
0.86
SVDQuant ( Li et al., 2025a )
23.71
0.7125
0.2906
0.2101
0.4931
63.50
0.3441
0.5342
0.95
Table 1: Quantitative results on synthetic and real-world datasets. WAN stands for our one-step inference version for VSR tasks based on training settings of DOVE ( Chen et al., 2025 ) , with pretrained WAN2.1 ( Wan et al., 2025 ) as our FP backbone components. Red and Blue texts represent the best and second-best scores, respectively.
Method
Bits
Params / M ( ↓ Ratio)
Ops / G ( ↓ Ratio)
WAN2.1
W16A16
1419 ( ↓ 0%)
40110( ↓ 0%)
W8A8
766 ( ↓ 45.97%)
21284 ( ↓ 46.94%)
LSGQuant
W6A6
592 ( ↓ 58.23%)
16274 ( ↓ 59.43%)
W4A4
419 ( ↓ 70.49%)
11264 ( ↓ 71.92%)
Table 2: A comparison of Params, Ops, and compression ratios (DiT module only) across various quantization settings. Ops are computed using a latent input size of 16x9x90x158, which represents a 32x3x180x318 video in the UDM10 dataset.
Figure 5: Visual comparison on synthetic UDM10 ( Tao et al., 2017 ) , REDS30 ( Nah et al., 2019 ) and real-world MVSR4x ( Wang et al., 2023b ) datasets at 4-bit quantization. Our approach outperforms existing methods in the 4-bit setting.
Figure 6: Comparison of temporal consistency (by stacking the red line across frames). Our method achieves better spatial and temporal consistency than the compared approaches.
Table 3: Ablation study on different method settings about DRAQ, VOLTS, and QAO. Experiments are conducted on the dataset HQ-VSR ( Chen et al., 2025 ) with full-precision inference, and evaluated on the synthetic dataset UDM10 ( Tao et al., 2017 ) .
Diffusion model quantization compresses model weights and activations to reduce inference memory and latency while preserving generated image, video, and speech quality.
Video super-resolution (VSR) using large-scale Diffusion Transformer (DiT) priors achieves exceptional perceptual quality but is often impractical due to the quadratic computational cost of processing dense spatio-temporal token sequences. Existing efficiency-oriented methods risk irreversible detail loss and temporal flickering, a vulnerability especially pronounced in one-step diffusion models. To address this, we propose TRaM-VSR, a Token Routing and Merging framework for adaptive token allocation, leveraging both context-aware video priors and network-level priors. First, token importance is estimated by fusing motion-sensitive temporal cues with semantic text similarity, isolating dynamic objects and structural boundaries. Next, this importance is further calibrated and adjusted by an offline planner to guide routing across optimally grouped network blocks. Technically, within each routed group, structurally critical tokens are processed in a high-fidelity local stream, while less informative tokens are aggregated into a compact global stream, both modulated by network depth and aligned with the multigranular nature of diffusion models. Extensive experiments show that TRaM-VSR accelerates inference significantly while preserving state-of-the-art reconstruction quality and robust temporal consistency. The code is available at https://github.com/Ree1s/TRaM-VSR.
Sicheng Gao, Zhuyun Zhou, Yixuan Liu +3
Advanced Micro Devices Inc. · Computer Vision Lab, CAIDAS & IFI, University of Würzburg
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.
Zhengyao Lv, Menghan Xia, Xintao Wang +1
The University of Hong Kong · Huazhong University of Science and Technology · Kling Team, Kuaishou Technology
Video diffusion models (VDMs) have achieved impressive progress in text-to-video generation, but their high memory and computational costs hinder practical deployment. Quantization-aware training (QAT) is an effective solution for compressing and accelerating advanced generative models without runtime overhead at inference. However, existing QAT methods suffer from a distinctive challenge in VDMs: while they often preserve prompt semantics, global layout, and coarse motion, the quantized model severely degrades visual details, texture fidelity, and sharpness. In this paper, we trace this degradation to the timestep-agnostic design of conventional quantization pipelines, which overlooks the stage-wise functionality of video denoising. In VDMs, early denoising steps mainly establish global structure and motion, whereas middle and late steps refine local appearance and high-frequency details. Based on this insight, we propose DSAQuant, a Denoising-Stage-Aligned Quantization-aware training framework for VDMs. During training, Denoising-Stage Oriented Supervision preserves teacher distillation in early steps for stable structure planning, while shifting later steps toward target-driven optimization to enhance detail reconstruction. During inference, Denoising-Stage Gated Guidance disables CFG in the final denoising steps to prevent it from amplifying quantization-induced errors into high-frequency artifacts. Extensive experiments on the Wan and CogVideoX families under W4A4 and W3A3 settings show that DSAQuant consistently outperforms the SOTA QAT baseline, improving the VBench average score by up to 6.60 under aggressive W3A3 quantization while preserving strong text-video alignment. These results demonstrate that effective VDM quantization requires not only reducing quantization error, but also aligning quantization training and inference with the stage-wise nature of video diffusion.