MDFI: A Multi-Domain Features Integration for Compressed Video Quality Enhancement

Sang NguyenQuang; Hieu Bui Minh; Dang BuiDinh; Xiem HoangVan

eess.IVAug 21, 2026

MDFI: A Multi-Domain Features Integration for Compressed Video Quality Enhancement

Authors: Sang NguyenQuang, Hieu Bui Minh, Dang BuiDinh, Xiem HoangVan

Organizations: Department of Computer Science, National Yang Ming Chiao Tung University, Hsinchu 30010, Taiwan · Faculty of Electronics and Telecommunications, VNU-University of Engineering and Technology, Vietnam National University, Hanoi 100000, Vietnam

Abstract

The latest video coding standard, H.266/VVC, has demonstrated significant improvements in compression efficiency compared to H.265/HEVC. Despite its advanced coding techniques, H.266/VVC still faces challenges in meeting the increasing demand for higher perceptual quality and enhanced compression performance. To address these limitations, we propose MDFI (Multi-Domain Features Integration), a compressed video quality enhancement approach that features a novel Frame-Prediction Feature Transform (FPFT) module to process prediction information. Moreover, MDFI integrates a multi-domain feature fusion strategy that effectively combines spatiotemporal characteristics, cross-frequency representations, and compressed-domain prediction information to enhance decoded video quality. Additionally, we introduce a comprehensive dataset that encompasses uncompressed video sequences, corresponding reconstructed versions at multiple QP levels, and predicted frames generated from H.266/VVC compressed bitstreams, providing essential resources for developing and benchmarking video enhancement approaches. Extensive experiments demonstrate that our MDFI approach achieves superior performance to state-of-the-art methods in both objective metrics and visual quality, effectively mitigating video compression artifacts. The code is available at: https://github.com/dangdinh17/MDFI.git.

Figures & tables

Fig. 1: PSNR comparison and exploited information in TGAF [ 38 ] , TVQE [ 35 ] , OVQE [ 22 ] and our MDFI.

Fig. 2: Overview of the MDFI architecture with the proposed Frame-Prediction Feature Transform. The full framework is illustrated in the gray background, while the SFT submodule is highlighted in the bottom-right corner.

Panel of Figure 3 — Fig. 3: Visualization of decoded information used in MDFI.

Dataset	MFQEv2 [ 14 ]	MDFI (Ours)
Total sequences	160	126
Training sequences	106	108
Validation sequences	36	$\times$
Testing sequences	18	18
Predicted sequences	$\times$	✓
Diversity	Interview, Sports, Natural, Urban	Interview, Sports, Natural, Urban

TABLE I: MFQEv2 dataset [ 14 ] versus our MDFI dataset.

QP	Class	Sequence	TGAF [ 38 ]		TVQE [ 35 ]		OVQE [ 22 ]		Wang et al [ 25 ]		MDFI (Ours)
QP	Class	Sequence	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM
37	A (2560 $\times$ 1600)	Traffic	0.67	0.94	0.67	0.97	0.87	1.21	0.90	1.25	1.04	1.40
	A (2560 $\times$ 1600)	PeopleOnStreet	0.96	1.80	0.90	1.72	1.02	1.91	1.06	2.01	1.21	2.22
	B (1920 $\times$ 1080)	Kimono	0.72	1.27	0.70	1.17	0.88	1.52	0.87	1.52	1.07	1.81
		ParkScene	0.59	1.47	0.53	1.34	0.73	1.85	0.76	1.97	0.89	2.18
		Cactus	0.51	1.00	0.51	0.93	0.61	1.22	0.62	1.21	0.71	1.35

TABLE II: Overall comparison for

\Delta

PSNR (dB) and

\Delta

SSIM (

\times 10^{-2}

) over standard test sequences at four QPs

Panel of Figure 6 — Fig. 4: PSNR comparison across frames on BasketballDrill (top) and BQMall (bottom) sequences at QP = 37.

Sequence		TGAF [ 38 ]	TVQE [ 35 ]	OVQE [ 22 ]	Wang et al [ 25 ]	MDFI
A	Traffic	-18.31	-18.59	-25.15	-25.50	-29.23
A	PeopleOnStreet	-13.39	-12.48	-15.00	-15.82	-18.91
B	Kimono	-17.31	-15.06	-20.53	-19.86	-24.83
	ParkScene	-20.60	-17.34	-26.77	-27.16	-30.80
	Cactus	-15.12	-14.08	-19.09	-19.11	-22.45
	BQTerrace	-12.71	-12.39	-18.64	-18.81	-23.46

TABLE III: BD-rate

(\%)

with H.266/VVC as the anchor.

Panel of Figure 8 — Fig. 5: Visual comparison of results between experimented methods. The sources of the video frames in the visualization results are: BasketballPass (416 × 240), BQMall (832 × 480).

Fig. 6: Rate-Distortion performance comparison on Traffic, BQSquare, RaceHorses, FourPeople.

Fig. 7: Trade-off analysis between quality improvement and model complexity.

Model	Parameters (M)	FLOPs (T)	$\Delta$ PSNR (dB)	BD-rate (%)
TGAF [ 38 ]	1.39	1.32	0.62	-16.82
TVQE [ 35 ]	2.08	0.12	0.60	-14.95
OVQE [ 22 ]	3.06	2.13	0.74	-21.03
Wang et al [ 25 ]	2.51	4.42	0.75	-21.35
MDFI (Ours)	3.84	2.85	0.85	-24.66

TABLE IV: Parameters, FLOPs,

\Delta

PSNR, and BD-rate comparisons of different methods at QP = 37.

Model	Configuration	Parameters (M)	FLOPs (T)	Time (ms)	$\Delta$ PSNR (dB)	Ratio
MDFI	out_nc = 64, $N=16$	3.84	2.85	724	0.85	3.35
MDFI_S1	out_nc = 64, $N=1$	3.44	2.48	550	0.83	2.99
MDFI_S2	out_nc = 32, $N=16$	2.22	1.61	501	0.74	2.17
MDFI_S3	out_nc = 32, $N=1$	1.82	1.24	326	0.62	2.00
OVQE	Baseline	3.06	2.13	475	0.74	2.88

TABLE V: Ablation study under different out_nc and

N

settings of MDFI, compared with baseline OVQE at QP=37.

FPFT	STFF	GMFF	FQE	$\Delta$ PSNR	$\Delta$ SSIM
			✓	0.35	0.74
		✓	✓	0.46	1.00
	✓	✓	✓	0.74	1.49
✓	✓	✓	✓	0.85	1.68

TABLE VI: Impact of different MDFI components on performance in terms of

\Delta

PSNR and

\Delta

SSIM at QP=37.

Model	$\Delta$ PSNR $\uparrow$	$\Delta$ SSIM $\uparrow$	$\Delta$ VMAF $\uparrow$	$\Delta$ LPIPS $\downarrow$
TGAF [11]	0.62	1.23	2.33	-0.63
TVQE [12]	0.60	1.19	2.86	-0.49
OVQE [13]	0.74	1.49	2.96	-1.01
Wang et al [14]	0.75	1.48	3.13	-0.79
MDFI (Our)	0.85	1.68	3.79	-1.17

TABLE VII: Overall evaluation for

\Delta

PSNR (dB),

\Delta

SSIM (

\times 10^{-2}

) ,

\Delta

VMAF and

\Delta

LPIPS (

\times 10^{-2}

) over standard test sequences at QP=37.

Video Compression

Video compression reduces the bits needed to store or transmit moving images while balancing visual quality, computational cost, and downstream usefulness.

See all 73 papers

DiffCVE: Diffusion-based Compressed Video Enhancement
Jul 8, 2026Wenqiang Xiao, Wenzhuo Ma, Junxi Zhang +1Video Diffusion Models Video Compression
School of Remote Sensing and Information Engineering, Wuhan University
Efficient Video Diffusion with Sparse Information Transmission for Video Compression
Mar 19, 2026Mingde Zhou, Zheng Chen, Yulun ZhangGenerative Video Compression Video Diffusion Models
Shanghai Jiao Tong University
FDIM: A Feature-distance-based Generic Video Quality Metric for Versatile Codecs
Apr 27, 2026Jiayi Wang, Lichun Zhang, Xiaoqi Zhuang +3Video Quality Assessment Video Compression
College of Information Science and Electronic Engineering, Zhejiang University, and also with the Zhejiang Key Laboratory of Multimodal Communication Networks and Intelligent Information Processing, Hangzhou 310058, China · Central Media Technology Institute, Huawei, Hangzhou 310058, China

Jul 8, 2026J/K move · Enter open · S save

DiffCVE: Diffusion-based Compressed Video Enhancement

Wenqiang Xiao, Wenzhuo Ma, Junxi Zhang, Zhenzhong Chen

School of Remote Sensing and Information Engineering, Wuhan University

Video Diffusion Models Video Compression

Open paper AskPDF arXiv

Perceptual quality enhancement of severely compressed videos remains challenging due to complex artifact patterns and substantial information loss. Recent diffusion models have demonstrated strong generative capability for visual restoration, but directly applying them to compressed video often ignores compression degradation characteristics and may introduce structure-inconsistent hallucinations. To address this issue, this paper presents a diffusion-based compressed video enhancement method, named DiffCVE. Coding Prior-enhanced Dual Conditioning (CPDC) branches are designed to jointly model compressed video and coding prior conditions, where coding priors including residuals and motion vectors provide complementary structural and motion guidance during the diffusion denoising process. To make the diffusion process aware of compression severity, a Compression Degradation Semantic Prompting (CDSP) mechanism is introduced to leverage QP-conditioned textual prompts together with LoRA fine-tuning. In addition, a Coding Prior-guided Weighted Fusion (CPWF) module is incorporated into the VAE decoder to fuse VAE encoder and coding prior encoder features with QP-predicted weights. Extensive experiments demonstrate the effectiveness of the proposed method in improving perceptual quality, especially under severe compression settings. The project page with enhanced video demonstrations is available at https://wqmaker.github.io/projects/DiffCVE/.

Abstract

Video Compression

Explore similar work

DiffCVE: Diffusion-based Compressed Video Enhancement

Efficient Video Diffusion with Sparse Information Transmission for Video Compression

FDIM: A Feature-distance-based Generic Video Quality Metric for Versatile Codecs