Real-time image denoising is essential for modern mobile photography but remains challenging due to the strict latency and power constraints of edge devices. This paper presents FastSHADE (Fast Self-augmented Hierarchical Asymmetric Denoising), a lightweight U-Net-style network tailored for real-time, high-fidelity restoration on mobile GPUs. Our method features a multi-stage architecture incorporating a novel Asymmetric Frequency Denoising Block (AFDB) that decouples spatial structure extraction from high-frequency noise suppression to maximize efficiency, and a Spatially Gated Upsampler (SGU) that optimizes high-resolution skip connection fusion. To address generalization, we introduce an efficient Noise Shifting Self-Augmentation strategy that enhances data diversity without inducing domain shifts. Evaluations on the MAI2021 benchmark demonstrate that our scalable model family establishes a highly efficient speed-fidelity trade-off. Our base FastSHADE-M variant maintains real-time latency (<50 ms on an Adreno 840 GPU) while preserving structural integrity, and our scaled-up FastSHADE-XL establishes a new state-of-the-art for overall image quality, achieving 37.94 dB PSNR.
Figures & tables
Figure 1: Latency-Quality Pareto Frontier. The FastSHADE model family establishes a highly efficient trade-off between on-device inference time (measured in FP16 on a Qualcomm Adreno 840 GPU) and pixel-level restoration fidelity (PSNR). While ultra-lightweight methods like TLG operate faster, they suffer a noticeable degradation in image quality. In contrast, our scalable architecture significantly expands the state-of-the-art efficiency frontier, pushing the boundaries of visual fidelity while offering highly competitive, low-latency variants for real-world mobile denoising.
Figure 2: Proposed model architecture
Figure 3: Frequency domain analysis of the internal features within the AFDB Block, averaged over 50 validation images. (Left, Center) The 2D FFT magnitude spectra exhibit a broader energy distribution in the 5×5 path compared to the 3×3 path. (Right) In the normalized radially averaged power spectrum, the spatial path (blue) retains significant energy in the low-to-mid spatial frequencies, indicating the preservation of structural details and edges. In contrast, the HF noise path (red) has a flat spectral response characteristic of isotropic, unstructured noise extraction.
Figure 4: Qualitative results of FastSHADE model predictions.
Method
SSIM ↑
PSNR ↑
t (ms) ↓
Score ↑
TLG
0.9129
37.31
28.0
13.72
SplitterNet [ 6 ]
-
37.62
66.1
8.93
TeamPAK
0.9098
37.58
127.0
4.40
Z6
0.9001
36.96
69.3
3.41
IVCL
0.9110
37.28
1961
0.19
FastSHADE-M
0.9136
37.52
48.1
10.68
Table 1: Performance metrics of FastSHADE compared to baselines and other methods. SSIM and PSNR values are reported for the MAI2021 test subset. Runtime, t , is reported for the Adreno 840. SSIM for SplitterNet was not reported because the test subset is not public.
Downscaling Layer
PSNR(dB) ↑
t (ms) ↓
Score ↑
Haar (frozen)
37.475
151
3.198
Haar (learnable)
37.492
151
3.275
3×3 Conv (stride 2)
37.581
141
3.967
2×2 Conv (stride 2)
37.580
136
4.108
Table 2: Ablation study on spatial downscaling strategies. Replacing the Haar wavelet transform and projection layer with a single learnable 2×2 strided convolution achieves the highest score.
Channel Split LF:HF
PSNR(dB) ↑
t (ms) ↓
Score ↑
1:0 (No Split)
36.56
144
3.773
1:1 (Symmetric)
37.58
142
3.934
1:3 (AFDB)
37.58
136
4.108
Table 3: Evaluation of channel splitting strategies in the AFDB. 1(LF):0(HF) routes all features through the 5×5 spatial branch; 1(LF):1(HF) represents a symmetric split; and 1(LF):3(HF) is our proposed asymmetric configuration (25% 5×5 spatial branch, 75% 3×3 noise). The asymmetric split matches the PSNR of the symmetric baseline but reduces inference time.
Upsampler Variant
PSNR(dB) ↑
t (ms) ↓
Score ↑
Concat + 1×1 Conv
37.40
257
1.690
SE Attn.
37.37
145
2.863
CBAM Attn.
37.43
141
3.205
Tconv 2×2 + Sigmoid
37.54
142
3.700
SGU (w/o Conv3×3dw )
37.55
136
3.919
SGU
37.58
136
4.108
Table 4: Ablation study on skip connection fusion strategies in the upsampling block: compare the proposed SGU against standard U-Net concatenation, channel/spatial attention (SE and CBAM), and a transposed convolution baseline.
Data augmentation
PSNR(dB) ↑
Baseline
36.80
+ basic
37.58
+ basic + noise swap
37.51
+ basic + our
37.63
Table 5: Data augmentation techniques ablation results. Baseline means no augmentations; + basic adds random rotations, flips, and artificial Gaussian noise; ( σ=25 ); + basic + noise swap includes basic augmentation and random noise profile swap between images in a batch; + basic + our adds the proposed Noise Shifting Self-Augmentation.
While deep-learning-based image restoration has achieved unprecedented fidelity, deployment on mobile Neural Processing Units (NPUs) remains bottlenecked by operator incompatibility and memory-access overhead. We propose an NPU-aware hardware-algorithm co-design approach for real-world image denoising on mobile NPUs. Our approach employs a high-capacity teacher to supervise a lightweight student network specifically designed to leverage the tiled-memory architectures of modern mobile SoCs. By prioritizing NPU-native primitives -- standard 3x3 convolutions, ReLU activations, and nearest-neighbor upsampling -- and employing a progressive context expansion strategy (up to 1024x1024 crops), the model achieves 37.66 dB PSNR / 0.9278 SSIM on the validation benchmark and 37.58 dB PSNR / 0.9098 SSIM on the held-out test benchmark at full resolution (2432x3200) in the Mobile AI 2026 challenge. Following the official challenge rules, the inference runtime is measured under a standardized Full HD (1088x1920) protocol, where it runs in 34.0 ms on the MediaTek Dimensity 9500 and 46.1 ms on the Qualcomm Snapdragon 8 Elite NPU. We further reveal an "Inference Inversion" effect, where strict adherence to NPU-compatible operations enables dedicated NPU execution up to 3.88x faster than the integrated mobile GPU. The 1.96M-parameter student recovers 99.8% of the teacher's restoration quality via high-alpha knowledge distillation (alpha = 0.9), achieving a 21.2x parameter reduction while closing the PSNR gap from 1.63 dB to only 0.05 dB. These results establish hardware-aware distillation as an effective strategy for unifying high-fidelity denoising with practical deployment across diverse mobile NPU architectures. The proposed lightweight student model (LiteDenoiseNet) and its training statistics are provided in the NN Dataset, available at https://github.com/ABrain-One/NN-Dataset.
Faraz Kayani, Sarmad Kayani, Asad Ahmed +2
Computer Vision Lab, CAIDAS & IFI, University of Würzburg, Germany
Mobile image denoising requires both good restoration quality and low computational cost. In addition, it's annoying to collect large-scale LQ-GT clean pairs. As a result, we propose LiteKD-Net, a lightweight knowledge-distilled network for mobile image denoising. First, a physics-guided noise simulation pipeline generates paired training data by adding pixel crosstalk compared with pipelines applied to cameras. Next, we adapt the Real-ESRGAN to identity-resolution denoising and construct a lightweight Student using Lite-RRDB blocks based on depthwise separable convolutions. Third, feature-level knowledge distillation is applied to transfer the Teacher's restoration capability to the Student without introducing additional inference cost. Experiments on real-world datasets show that our model reaches great reduction in runtime and increase in the inference rate with good restoration quality. Our model also reaches the best in all metrics compared with SwinIR. These results indicate that LiteKD-Net provides a great trade-off between restoration quality and computational efficiency.
Zhou Zhiyi
Global College Shanghai Jiaotong University Shanghai, China
We propose a modular framework for hybrid image restoration that integrates transformer and state-space model (SSM) blocks with a focus on improving runtime efficiency on edge hardware. While transformers provide strong global modeling through self-attention, their attention kernels incur substantial latency on mobile devices, especially for high-resolution inputs. In contrast, SSMs such as Mamba offer lineartime sequence modeling with lower runtime overhead but may underperform on fine grained restoration tasks. To balance accuracy and efficiency, we train lightweight SSM blocks as feature-distilled surrogates of transformer blocks and use them to construct hybrid U-Net-style architectures. To automatically discover effective block combinations, we introduce Efficient Network Search (ENS), a multi-objective search strategy that selects task-specific hybrid configurations from pre-aligned components. ENS optimizes restoration quality while penalizing transformer usage, serving as a lightweight proxy for latency and enabling architecture discovery without repeated hardware profiling. On a Snapdragon 8 Elite CPU, the Restormer baseline requires 10119.52 ms for inference. In contrast, ENS-discovered hybrids significantly reduce runtime: ENS-Deblurring runs in 2973 ms (3.4x faster), ENS-Deraining in 5816 ms (1.74x faster), and ENS-Denoising in 8666 ms (1.17x faster), while maintaining competitive restoration quality.