UHD-GPGNet: UHD Video Denoising via Gaussian-Process-Guided Local Spatio-Temporal Modeling

Weiyuan He; Chen Wu; Pengwen Dai; Wei Wang; Dianjie Lu; Guijuan Zhang; Linwei Fan; Yongzhen Wang

cs.CVApr 13, 2026

UHD-GPGNet: UHD Video Denoising via Gaussian-Process-Guided Local Spatio-Temporal Modeling

Authors: Weiyuan He, Chen Wu, Pengwen Dai, Wei Wang, Dianjie Lu, Guijuan Zhang, Linwei Fan, Yongzhen Wang,

Organizations: Qingdao University, Qingdao, China · National University of Defense Technology, Changsha, China · SUN YAT-SEN University, Guangzhou, China · Shandong Normal University, Jinan, China ·

Abstract

Ultra-high-definition (UHD) video denoising requires simultaneously suppressing complex spatio-temporal degradations, preserving fine textures and chromatic stability, and maintaining efficient full-resolution 4K deployment. In this paper, we propose UHD-GPGNet, a Gaussian-process-guided local spatio-temporal denoising framework that addresses these requirements jointly. Rather than relying on implicit feature learning alone, the method estimates sparse GP posterior statistics over compact spatio-temporal descriptors to explicitly characterize local degradation response and uncertainty, which then guide adaptive temporal-detail fusion. A structure-color collaborative reconstruction head decouples luminance, chroma, and high-frequency correction, while a heteroscedastic objective and overlap-tiled inference further stabilize optimization and enable memory-bounded 4K deployment. Experiments on UVG and RealisVideo-4K show that UHD-GPGNet achieves competitive restoration fidelity with substantially fewer parameters than existing methods, enables real-time full-resolution 4K inference with significant speedup over the closest quality competitor, and maintains robust performance across a multi-level mixed-degradation schedule.A real-world study on phone-captured 4K video further confirms that the model, trained entirely on synthetic degradation, generalizes to unseen real sensor noise and improves downstream object detection under challenging conditions.

Figures & tables

Figure 1: Full-resolution 4K speed–quality–memory landscape. UHD-GPGNet ( $\bullet$ ) uniquely occupying the upper-right region. Smaller and higher-right is better.

Figure 2: Overall architecture of the proposed UHD video denoiser, including the Y/C/RGB stems, hierarchical multi-stage feature backbone, sparse GP-guided fusion blocks, structure-color reconstruction head, and full-resolution inference pipeline.

Figure 3: Sparse GP-guided fusion module. The module forms local descriptors, pools inducing tokens, predicts inducing-level mean and variance statistics, produces local posterior mean and uncertainty maps, and converts them into fusion gates that balance temporal aggregation against detail-preserving bypass features.

Method	PSNR $\uparrow$	SSIM $\uparrow$	Params(M) $\downarrow$	FLOPs(G) $\downarrow$
FastDVDnet	29.434	0.7675	2.479	334.4
ASwin	32.764	0.9688	8.910	841.0
BSVD	28.791	0.7178	9.816	3107.0
RVRT	28.751	0.7316	12.787	888.9
TAP †	31.061	0.8977	34.482	3183.2
UHD-GPGNet	32.759	0.9592	0.707	347.3

Table 1: Main comparison on UVG. FLOPs are the estimated 4K equivalent (

640\!\times\!640

per-tile cost

\times

tile count); all methods use matched tiled inference.

Figure 4: Main qualitative comparison on UVG with one full 4K frame and enlarged crops emphasizing thin structures, repeated textures, and chroma-sensitive regions.

Component	$\sigma{=}1.0$	$\sigma{=}2.0$	$\sigma{=}3.0$
Exposure	0.95–1.05	0.90–1.10	0.85–1.15
Sensor+ grain	shot 1.5e-3–1.2e-2 read 4e-4–4e-3 $g_{\sigma}$ 8e-4–8e-3	shot 3e-3–2.4e-2 read 8e-4–8e-3 $g_{\sigma}$ 1.6e-3–1.6e-2	shot 4.5e-3–3.6e-2 read 1.2e-3–1.2e-2 $g_{\sigma}$ 2.4e-3–2.4e-2
Flicker	$p{=}0.20$ ; $\rho{=}$ 0.85–0.97 std ${=}$ (0.02, 0.003)	$p{=}0.40$ ; $\rho{=}$ 0.85–0.97 std ${=}$ (0.04, 0.006)	$p{=}0.60$ ; $\rho{=}$ 0.85–0.97 std ${=}$ (0.06, 0.009)
Chroma	Cb/Cr std ${=}$ 0.002	Cb/Cr std ${=}$ 0.004	Cb/Cr std ${=}$ 0.006
H.264	CRF ${=}$ 28	CRF ${=}$ 32	CRF ${=}$ 36

Table 2: Mixed-degradation protocol at three anchor settings;

\sigma{=}1.5

2.5

follow the same monotonic scaling.

Method	$\sigma=1.0$	$\sigma=2.0$	$\sigma=3.0$	Avg. SSIM	Drop $1\!\rightarrow\!3$
FastDVDnet	33.43	29.17	26.58	0.768	6.85
ASwin	36.23	32.37	29.50	0.962	6.73
BSVD	32.29	28.53	26.18	0.718	6.11
RVRT	32.43	28.52	27.14	0.734	5.29
TAP †	35.30	31.02	28.33	0.910	6.97
UHD-GPGNet	36.43	32.14	30.22	0.951	6.21

Table 3: Cross-scene robustness on RealisVideo-4K under five-point mixed degradation. PSNR at anchor settings

\sigma{=}1.0

2.0

3.0

; average SSIM over all five points; PSNR drop from

\sigma{=}1.0

3.0

Figure 5: RealisVideo-4K robustness curves under controlled mixed degradations at $\sigma=1.0$ , $1.5$ , $2.0$ , $2.5$ , and $3.0$ .

Figure 6: Representative 4K crop visualization on RealisVideo-4K at $\sigma=1.0$ , $2.0$ , and $3.0$

Method	Peak Mem. (GiB) $\downarrow$	FPS $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$
FastDVDnet	1.32	62.52	30.250	0.7714
ASwin	7.31	0.95	34.177	0.9631
BSVD	3.29	12.95	29.640	0.7279
RVRT	5.20	2.04	29.485	0.7360
TAP †	4.94	3.33	32.496	0.9107
UHD-GPGNet	3.95	28.06	34.213	0.9639

Table 4: Full-resolution 4K deployment on RealisVideo-4K (

3840{\times}2160

) with synchronized measurement.

Variant	Det. gate	GP	Hetero	Y/C	HF	PSNR $\uparrow$	SSIM $\uparrow$
Backbone	$\times$	$\times$	$\times$	$\times$	$\times$	31.78	0.881
+ Det. gate	$\checkmark$	$\times$	$\times$	$\times$	$\times$	31.95	0.885
+ Attn+Var	$\times$	attn	$\times$	$\times$	$\times$	32.11	0.891
+ Attn+Var+P	$\times$	attn+p	$\times$	$\times$	$\times$	32.15	0.893
+ Sparse GP	$\times$	$\checkmark$	$\times$	$\times$	$\times$	32.32	0.898
+ Hetero obj.	$\times$	$\checkmark$	$\checkmark$	$\times$	$\times$	32.41	0.902

Table 5: Component ablation. “Attn+Var” replaces the GP kernel with dot-product attention while retaining variance prediction. “Attn+Var+P” further adds three matched learnable scalars (temperature, output scale, temporal weight) to control for parameter-count differences with the GP kernel (

\alpha

\ell

\gamma

Setting	Err–Unc Corr. $\uparrow$	Chroma Flicker $\downarrow$	CbCr PSNR $\uparrow$	Seam Score $\downarrow$	Ref. PSNR $\uparrow$
No GP	0.52	0.0234	37.19	0.0121	32.18
No hetero	0.59	0.0198	37.85	0.0113	32.29
RGB-only	0.61	0.0231	36.87	0.0107	32.21
No overlap	0.61	0.0185	37.96	0.0292	32.44
Full model	0.75	0.0125	39.12	0.0062	32.60

Table 6: Mechanism validation of uncertainty quality, chroma stability, chroma fidelity, and tiled deployment behavior.

Figure 7: Mechanism visualization: restored output, residual error, GP log-variance, predicted log-variance, and stage-2/3 fusion gates on a flare- and edge-contaminated region.

	Avg Det $\uparrow$	Avg Conf $\uparrow$	Temp Cons $\uparrow$	NIQE $\downarrow$	BRISQUE $\downarrow$
Raw input	6.2	0.414	0.863	3.66	19.8
Denoised	6.5	0.423	0.878	3.55	19.0
$\Delta$	+0.3	+0.009	+0.015	$-$ 0.11	$-$ 0.8

Table 7: Real-world phone-captured 4K evaluation (averaged across scenes). The frozen pretrained model is applied without retraining. Detection uses YOLOv8; quality uses no-reference metrics (lower = better).

Figure 8: Real-world downstream detection on phone-captured 4K video. Top : outdoor dusk scene—denoising recovers distant targets and increases detection confidence. Bottom : indoor low-light scene—object boundaries become sharper and detection boxes stabilize. The model is applied without retraining on real noise.

Image Restoration

Image restoration uses computational methods to recover useful visual content from noisy, blurred, low-resolution, or otherwise degraded images.

See all 372 papers

From Zero to Detail: A Progressive Spectral Decoupling Paradigm for UHD Image Restoration with New Benchmark
Apr 17, 2026Chen Zhao, Yunzhe Xu, Zhizhou Chen +5Frequency-Domain Image Processing Image Restoration
State Key Laboratory of Novel Software Technology, Nanjing University, Nanjing 210023, China · School of Computer Science and Technology, Dalian University of Technology, Dalian 116023, China · Department of Computer Science and Engineering, Michigan State University, East Lansing, MI 48824, U.S.A.
DTG-Restore: Training-Free Diffusion Refinement for Generative Video Super-Resolution
May 28, 2026Hidir Yesiltepe, Koutilya PNVR, Gaurav Pathak +4Diffusion Model Guidance Video Diffusion Models
1Virginia Tech · 2Adobe
LoopMoEVR: Loop-Based Degradation-Aware Mixture-of-Experts for Unified UHD Video Restoration
Oct 4, 2026Yucheng Xin, Runci Bai, Yongcong Wang +5Mixture of Experts Image Restoration
Shandong Normal University · China Academy of Information and Communications Technology · Central South University +4

Apr 17, 2026J/K move · Enter open · S save

From Zero to Detail: A Progressive Spectral Decoupling Paradigm for UHD Image Restoration with New Benchmark

Chen Zhao, Yunzhe Xu, Zhizhou Chen, Enxuan Gu+4

State Key Laboratory of Novel Software Technology, Nanjing University, Nanjing 210023, China · School of Computer Science and Technology, Dalian University of Technology, Dalian 116023, China · Department of Computer Science and Engineering, Michigan State University, East Lansing, MI 48824, U.S.A.

Frequency-Domain Image Processing Image Restoration Image Super-Resolution

Open paper AskPDF arXiv

Ultra-high-definition (UHD) image restoration poses unique challenges due to the high spatial resolution, diverse content, and fine-grained structures present in UHD images. To address these issues, we introduce a progressive spectral decomposition for the restoration process, decomposing it into three stages: zero-frequency enhancement, low-frequency restoration, and high-frequency refinement. Based on this formulation, we propose a novel framework, ERR, which integrates three cooperative sub-networks: the zero-frequency enhancer (ZFE), the low-frequency restorer (LFR), and the high-frequency refiner (HFR). The ZFE incorporates global priors to learn holistic mappings, the LFR reconstructs the main content by focusing on coarse-scale information, and the HFR adopts our proposed frequency-windowed Kolmogorov-Arnold Network (FW-KAN) to recover fine textures and intricate details for high-fidelity restoration. To further advance research in UHD image restoration, we also construct a large-scale, high-quality benchmark dataset, LSUHDIR, comprising 82{,}126 UHD images with diverse scenes and rich content. Our proposed methods demonstrate superior performance across a range of UHD image restoration tasks, and extensive ablation studies confirm the contribution and necessity of each module. Project page: https://github.com/NJU-PCALab/ERR.

Abstract

Image Restoration

Explore similar work

From Zero to Detail: A Progressive Spectral Decoupling Paradigm for UHD Image Restoration with New Benchmark

DTG-Restore: Training-Free Diffusion Refinement for Generative Video Super-Resolution

LoopMoEVR: Loop-Based Degradation-Aware Mixture-of-Experts for Unified UHD Video Restoration