Real-world video denoising for communication systems must handle mixed degradations such as sensor noise, compression artifacts, and motion-induced inconsistency while remaining deployable under tight latency budgets. Purely deep models often achieve high accuracy but come with significant computational cost, whereas classical filters are robust and efficient but insufficiently adaptive. We present FilterFusion, a prior-guided multi-frame denoising framework that combines temporally-aware learned features with an interpretable bank of denoising branches. The model predicts a low-cost guidance signal and uses it to adaptively mix a compact set of denoising responses at the original resolution. The lightweight design uses only 109K parameters and explicitly injects temporal variance and spatial variance cues into the low-resolution controller. On CRVD (real sensor noise), it reaches 34.72 dB PSNR and 0.9330 SSIM with only 0.18 GFLOPs (network-only). Because the filter-bank operations are simple and regular (Gaussian blur, median, weighted sum), a straightforward compiled C implementation of the synthesis path achieves 17.9ms per frame (55.9 FPS) on a single CPU thread at 1080p---exceeding real-time and 122--252 times faster than lightweight CNN baselines that require over 2 seconds per frame. The results demonstrate a strong quality-efficiency-speed trade-off and a practical route to real-time denoising through low-resolution control and explicit filter-based synthesis.
Figures & tables
Fig. 1: Multi-dimensional trade-off. Each axis is normalized to [0,1] : Quality (PSNR, [22,40] dB), Lightness ( 1−params/106 ; classical = max), Structure (SSIM, [0.6,1.0] ), Speed ( log FPS), Efficiency ( log1/GFLOPs ). The dashed green arc marks the real-time threshold (30 FPS). FilterFusion (red) is the only method that simultaneously exceeds the real-time boundary and surpasses the classical quality ceiling.
Fig. 2: Overview of FilterFusion in luma plane. A compact low-resolution predictor estimates pixel-wise fusion weights from the reference frame Yref , temporal mean Yˉ , and variance guidance (Vs,Vt) , while a full-resolution filter bank generates multiple candidate responses. The final prediction is obtained by adaptively fusing these responses using the predicted weights. For the chroma planes, the same filter-bank synthesis is applied while reusing the weights predicted from the luma plane, producing the denoised NV12 output with negligible additional cost.
Method
Params
PSNR ↑
SSIM ↑
Time (ms) ↓
GFLOPs ↓
Real-time methods ( ≥ 30 FPS)
Gaussian blur
–
32.54
0.8713
7 ⋆
∼ 0.06
Box filter
–
31.61
0.8822
9 ⋆
∼ 0.04
FilterFusion
109K
34.72
0.9330
17.9 ⋆‡
0.18 †
Median filter
–
32.48
0.8779
21 ⋆
∼ 0.16
Non-real-time methods ( < 30 FPS)
TABLE I: Comparison on CRVD at 1920×1080 , sorted by runtime within each group. All timings are on a single CPU thread (Intel i9-10900X) for our methods and baselines.
Fig. 3: Role of Vt gating on CRVD scene7/ISO12800. Inset (top-left) is the predicted Vt map (warm = high, cool = low). On high- Vt regions (red box), naive multi-frame averaging ghosts the KIPSTA logos while FilterFusion preserves edges. On low- Vt regions (green box), FilterFusion suppresses sensor noise without blurring the thin power lines.
#
Variant
PSNR ↑
SSIM ↑
(a)
Full model (FilterFusion)
34.7549
0.9353
Guidance signals
(b)
w/o temporal variance Vt
34.8418
0.9336
(c)
w/o spatial variance Vs
34.6080
0.9304
(d)
w/o neither Vt nor Vs
34.6354
0.9316
Filter-bank composition
TABLE II: Ablation study on CRVD. Each row disables or modifies one component relative to the full model. All variants are trained single-stage on the same data split as the full model in this table (no SRVD pre-training), with the same schedule (160K iterations, Adam, same LR milestones); they are used to attribute quality to specific design choices and are not directly comparable to the two-stage main-table numbers.
Fig. 4: Qualitative ablation on CRVD ISO 25600 (strong noise, high Vt ): (a) remains heavy residual noise; (b) noise is cleaned but visible ghosting along the ball’s moving edge; (c) clean and ghost-free; (d) ground truth.
Autoregressive video diffusion models have become a leading approach to real-time interactive video generation and video world modeling. However, existing methods typically allocate computation uniformly across space when generating the next video chunk, updating all patches with the same denoising schedule and number of function evaluations. This ignores the spatial heterogeneity of video content: static, simple, or recurring regions are often easier to denoise, while other regions require further refinement or additional context. A uniform budget thus leads to redundant computation on easy patches, whereas reducing the fixed step count can leave difficult patches insufficiently refined. We introduce \methodname{}, a general framework that dynamically allocates denoising steps to individual patches: easy patches stop early to improve efficiency, while difficult patches receive more refinement steps. This design accelerates generation with negligible quality degradation. Our method is compatible with most existing few-step autoregressive diffusion models, requiring only a lightweight routing module with modest training. Experiments demonstrate that \methodname{} reduces the total number of denoising steps applied across video patches by about 50% while maintaining visual quality comparable to the four-step baseline.
Zekai Gu, Peng Wang, Jiahao Lu +3
HKUST · VAST · Macau University of Science and Technology
A spatial-local image-denoising filter is proposed, and its performance metrics are evaluated in comparison with baseline filtering algorithms, including the median, adaptive median, Gaussian, bilateral, Wiener, anisotropic diffusion, and non-local means. The developed filter is based on aligning the intensity value of the central pixel in a 3x3 window with the statistical majority intensity of one of the two clusters formed by optimal Otsu's partitioning of a pixel set sorted by intensity and trimmed to seven elements. This is followed by a fuzzy fusion of the calculated value with the median intensity of the pixels within the window. The proposed filter demonstrates the highest robustness to variations in image noise levels, particularly when processing mixed noise consisting of salt-and-pepper impulse noise and additive Gaussian noise in various proportions
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.
Weiyuan He, Chen Wu, Pengwen Dai +6
Qingdao University, Qingdao, China · National University of Defense Technology, Changsha, China · SUN YAT-SEN University, Guangzhou, China +3