Decoder-Free Distillation for Quantized Image Restoration

S. M. A. Sharif; Abdur Rehman; Seongwan Kim; Jaeho Lee

cs.CVMar 10, 2026

Decoder-Free Distillation for Quantized Image Restoration

Authors: S. M. A. Sharif, Abdur Rehman, Seongwan Kim, Jaeho Lee

Organizations: Opt-AI Inc., Seoul, South Korea

Abstract

Quantization-Aware Training (QAT), combined with Knowledge Distillation (KD), holds immense promise for compressing models for edge deployment. However, joint optimization for precision-sensitive image restoration (IR) to recover visual quality from degraded images remains largely underexplored. Directly adapting QAT-KD to low-level vision reveals three critical bottlenecks: teacher-student capacity mismatch, spatial error amplification during decoder distillation, and an optimization "tug-of-war" between reconstruction and distillation losses caused by quantization noise. To tackle these, we introduce Quantization-aware Distilled Restoration (QDR), a framework for edge-deployed IR. QDR eliminates capacity mismatch via FP32 self-distillation and prevents error amplification through Decoder-Free Distillation (DFD), which corrects quantization errors strictly at the network bottleneck. To stabilize the optimization tug-of-war, we propose a Learnable Magnitude Reweighting (LMR) that dynamically balances competing gradients. Finally, we design an Edge-Friendly Model (EFM) featuring a lightweight Learnable Degradation Gating (LDG) to dynamically modulate spatial degradation localization. Extensive experiments across four IR tasks demonstrate that our Int8 model recovers 96.5% of FP32 performance, achieves 442 frames per second (FPS) on an NVIDIA Jetson Orin, and boosts downstream object detection by 16.3 mAP

Figures & tables

Panel of Figure 1 — Figure 1: Comparison of standard Distillation and our proposed Quantization-aware Distill Restoration (QDR) paradigm ( Left ). QDR enables high-quality restoration across diverse degradations (low-light, deraining, denoising, dehazing) ( Middle ). Our method recovers $\sim$ 96.5% of FP32 and significantly faster FPS (442) on the edge board ( Right ).

Figure 2: Overview of the proposed QDR framework, which consists of two components: DFD and LMR. The DFD architecture ensures effective knowledge distillation under quantization, while the LMR module dynamically reweights the magnitudes of reconstruction and distillation losses to improve training stability and performance.

Panel of Figure 3 — Figure 3: Activation histograms at the bottleneck and decoder layers, illustrating the distribution differences between the FP32 and quantized models across different architectures.

Figure 4: Architecture of our proposed Learnable Degradation Gating module containing two trainable scalers $[\alpha_{\ell},\alpha_{d}]$ for controlling the fusion of Degradation importance map with residual features.

Type	Method	Complexity				Denoise (SIDD)			Lowlight (LoL-v1)			Derain (Rain100H)			Dehaze (SOTS)			Average
Type	Method	Flops	Params	FPS 3090	FPS ORIN	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Trans.	Uformer [ 34 ]	10.87	5.29	52	-	37.20	0.879	0.210	22.83	0.870	0.108	26.29	0.829	0.118	25.98	0.948	0.027	28.08	0.882	0.116
	Restormer [ 40 ]	141.24	26.13	9	-	40.02	0.960	0.149	20.41	0.806	0.128	31.46	0.904	0.081	30.87	0.969	0.018	30.69	0.910	0.094
	MIRNetv2 [ 42 ]	140.92	5.18	17	6	39.84	0.959	0.152	24.74	0.851	0.071	26.35	0.824	0.115	26.26	0.956	0.035	29.30	0.898	0.093
CNN	U-Net [ 31 ]	15.53	9.50	207	144	27.33	0.494	0.559	17.18	0.722	0.305	19.97	0.598	0.375	24.41	0.932	0.046	22.22	0.687	0.321
	DnCNN [ 43 ]	36.73	0.67	172	89	31.53	0.933	0.249	18.67	0.732	0.246	24.28	0.796	0.176	25.18	0.928	0.060	24.92	0.847	0.183

Table 1: Quantitative comparison with state-of-the-art methods. Red/Bold and Blue/Underlined indicates the best and the second best among CNNs respectively.

Figure 5: Qualitative results on four restoration tasks: denoising, LLIE, deraining, and dehazing on full-precision settings.

Method	Denoise (SSID)			Lowlight (LoL-v1)			Derain (Rain100H)			Dehaze (SOTS)			Average
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
FP32 (Reference)	37.22	0.877	0.227	22.43	0.799	0.174	30.03	0.886	0.111	28.10	0.961	0.018	29.64	0.900	0.110
PTQ	29.31	0.584	0.491	18.86	0.579	0.398	28.10	0.814	0.165	26.85	0.919	0.037	25.78	0.724	0.273
QAT	33.09	0.729	0.491	21.76	0.749	0.233	28.36	0.823	0.158	27.19	0.926	0.033	27.60	0.809	0.222
QAT + KD [ 11 ]	33.58	0.734	0.358	21.99	0.758	0.226	28.47	0.828	0.154	27.51	0.928	0.030	27.89	0.817	0.183
QAT + FAKD [ 39 ]	33.31	0.709	0.491	22.10	0.762	0.223	28.44	0.827	0.155	27.86	0.931	0.030	27.93	0.816	0.212

Table 2: Quantitative comparison of INT8 quantization methods. We compare various quantization approaches on four restoration tasks. Red/Bold indicates the best while Blue/Underlined indicates the second best.

Figure 6: Qualitative results for quantized image restoration (8-bit): denoising, LLIE, deraining, and dehazing.

Model

Lowlight (LoL-v1)

Dehazing (SOTS)

Deraining (Rain100H)

2-bit

4-bit

8-bit

FP32

2-bit

4-bit

8-bit

FP32

2-bit

4-bit

8-bit

FP32

PTQ

DFD

PTQ

DFD

PTQ

DFD

PTQ

DFD

PTQ

DFD

PTQ

DFD

PTQ

DFD

PTQ

DFD

PTQ

DFD

Opt-LLIE [ 1 ]

9.30

9.80 +0.50

10.64

18.20 +7.56

20.04

21.20 +1.16

20.10

14.01

14.30 +0.29

19.27

22.69 +3.42

25.10

25.47 +0.37

25.58

11.93

10.70 -1.23

18.22

22.52 +4.30

26.97

27.18 +0.21

28.06

Ours

5.82

7.86 +2.04

9.91

13.47 +3.56

21.10

22.36 +1.26

22.43

17.71

18.20 +0.49

16.97

23.80 +6.83

26.50

28.72 +1.92

28.10

14.34

15.90 +1.56

11.34

18.47 +7.13

28.10

28.95 +0.85

30.03

Table 3: Quantitative comparison of proposed QDR vs. PTQ across 2-, 4-, and 8-bit widths. Our method consistently achieves higher PSNR (

\uparrow

) on all tasks.

Figure 7: Feature distribution alignment between different teachers and positions with the target FP32 Model.

Table 11

Figure 8: Comparison of GoR [ 30 ] with LMR during a training run.

Table 14

Figure S10 : Overview of proposed edge-friendly model.

Stage	Module	Channels	Spatial Size	Operation
Input proj.	Conv $3\times 3$	$3\to 64$	$H\times W$	Stem convolution
Encoder 1	ResBlock + SE	$64$	$H\times W$	Feature modulation
	MultiScaleDegMap	$64\to 1$	$H\times W$	Degradation map $G_{1}$
	Conv $3\times 3$ , stride 2	$64\to 96$	$\frac{H}{2}\times\frac{W}{2}$	Downsampling
Encoder 2	ResBlock + SE	$96$	$\frac{H}{2}\times\frac{W}{2}$	Feature modulation
	MultiScaleDegMap	$96\to 1$	$\frac{H}{2}\times\frac{W}{2}$	Degradation map $G_{2}$

Table S9 : EFM architecture overview.

H\times W

denotes the spatial resolution of the input image.

Panel of Figure 17 — Figure S11 : Hardware stability during inference. (a) clock speed and (b) Temperature across inference steps for different precision modes.

Figure S12 : Qualitative comparison of 8-bit image deraining on the Rain100H dataset.

Figure S13 : Qualitative comparison of 8-bit low-light enhancement on the LOL dataset.

Figure S14 : Qualitative comparison of 8-bit real-world denoising on the SIDD dataset.

Figure S15 : Qualitative comparison of 8-bit image dehazing on the SOTS dataset.

Figure S16 : Comparison of activation distributions between PTQ and our QDR method in the bottleneck layers.

Figure S17 : Comparison of activation distributions between PTQ and our QDR method in the decoder layers.

Figure S18 : Feature deviation comparison relative to the target FP model. We compare the standard PTQ model, the homogeneous (FP32) teacher-guided model, and the heterogeneous teacher-guided model.

Figure S19 : Visualization of degradation maps across different feature scales.

Method	MGD [ 38 ]	Perceptual [ 16 ]	L1	Ours (MSE)
PSNR (dB)	20.04	20.34	21.99	22.36

Table S10 : Ablation study on different loss functions for DFD.

Image Restoration

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

See all 357 papers

SPARK: Spatial Policy-driven Adaptive Reinforcement learning for Knowledge distillation
Jun 13, 2026Mohamed Jismy Aashik Rasool, Shabir Ahmad, Gisong Oh +1Quantization-Aware Training Image Restoration
Department of IT Convergence Engineering, Gachon University, Sujeong-gu, Seongnam-si, Gyeonggi-do 461-701, Republic of Korea
Universal Image Restoration via Internalized Chain-of-Thought Reasoning
Jun 16, 2026Yu Guo, Zhengru Fang, Shengfeng He +4All-in-One Image Restoration Image Restoration
Hong Kong JC Lab of Smart City and the Department of Computer Science, City University of Hong Kong · School of Computing and Information Systems, Singapore Management University · Computer Science and Information Engineering, National Taiwan University
Restore, Assess, Repeat: A Unified Framework for Iterative Image Restoration
Mar 27, 2026I-Hsiang Chen, Isma Hadji, Enrique Sanchez +5Image Restoration All-in-One Image Restoration
Samsung AI Center Cambridge · National Taiwan University · Technical University of Iasi +3

Jun 13, 2026J/K move · Enter open · S save

SPARK: Spatial Policy-driven Adaptive Reinforcement learning for Knowledge distillation

Mohamed Jismy Aashik Rasool, Shabir Ahmad, Gisong Oh, Teag Kuen Whangbo

Department of IT Convergence Engineering, Gachon University, Sujeong-gu, Seongnam-si, Gyeonggi-do 461-701, Republic of Korea

Quantization-Aware Training Image Restoration Knowledge Distillation Low-Bit Quantization Quantization-Aware Distillation

Open paper AskPDF arXiv

Low-bit quantization enables deployment of image restoration (IR) networks on resource-constrained devices, but introduces rounding noise that disproportionately degrades high-frequency regions such as edges and fine textures. Existing knowledge distillation (KD) methods apply distillation signals uniformly across all spatial locations, overlooking the varying reconstruction difficulty across image regions. To address this, we propose SPARK (Spatial Policy-driven Adaptive Reinforcement Learning for Knowledge Distillation), a framework that adaptively allocates distillation effort using a lightweight reinforcement learning (RL) policy network. At each training step, a difficulty feature extractor computes four signals, namely Laplacian variance, pixel variance, student reconstruction error, and teacher-student knowledge gap, which are fed into a compact policy CNN that produces a stochastic spatial weight map to modulate the KD loss during quantization-aware training (QAT). SPARK is IR task-agnostic, adds no inference cost, and integrates into any existing QAT pipeline without architectural changes. Experiments on benchmark datasets demonstrate that SPARK consistently outperforms PTQ, QAT, and state-of-the-art (SOTA) KD approaches across multiple student architectures, achieving reconstruction quality closest to the full-precision teacher under significant computational constraints.

Abstract

Image Restoration

Explore similar work

SPARK: Spatial Policy-driven Adaptive Reinforcement learning for Knowledge distillation

Universal Image Restoration via Internalized Chain-of-Thought Reasoning

Restore, Assess, Repeat: A Unified Framework for Iterative Image Restoration