IQ-LUT: interpolated and quantized LUT for efficient image super-resolution

Yuxuan Zhang; Zhikai Dong; Xinning Chai; Xiangyun Zhou; Yi Xu; Zhengxue Cheng; Li Song

cs.CVApr 8, 2026

IQ-LUT: interpolated and quantized LUT for efficient image super-resolution

Authors: Yuxuan Zhang, Zhikai Dong, Xinning Chai, Xiangyun Zhou, Yi Xu, Zhengxue Cheng, Li Song

Organizations: Shanghai Jiao Tong University · Rockchip Electronics Co., Ltd

Abstract

Lookup table (LUT) methods demonstrate considerable potential in accelerating image super-resolution inference. However, pursuing higher image quality through larger receptive fields and bit-depth triggers exponential growth in the LUT's index space, creating a storage bottleneck that limits deployment on resource-constrained devices. We introduce IQ-LUT, which achieves a reduction in LUT size while simultaneously enhancing super-resolution quality. First, we integrate interpolation and quantization into the single-input, multiple-output ECNN, which dramatically reduces the index space and thereby the overall LUT size. Second, the integration of residual learning mitigates the dependence on LUT bit-depth, which facilitates training stability and prioritizes the reconstruction of fine-grained details for superior visual quality. Finally, guided by knowledge distillation, our non-uniform quantization process optimizes the quantization levels, thereby reducing storage while also compensating for quantization loss. Extensive benchmarking demonstrates our approach substantially reduces storage costs (by up to 50x compared to ECNN) while achieving superior super-resolution quality.

Figures & tables

Panel of Figure 1 — Figure 1: Qualitative and quantitative comparison on Set5 for x4 SR. As can be seen from the left figure, our model IQ-L8C16 performs well on the boundary. As can be seen from the right figure, our three IQ-LUT models all achieve small LUT-Size and high PSNR.

Figure 2: The overview of our model.(a) represents the overall structure of IQ-LUT, (b) is the internal structure of IQ-Block, which is the core of our model. It consists of two parts: NUQD and DPFI. (c) is the internal structure of NUQD. The lower part is the distillation training process, the snowflakes on the left represent freezing, and the flames on the right represent trainable.

Model	Size(KB)	Set5	Set14	BSD100	Urban100	Manga109
		(PSNR $\uparrow$ SSIM $\uparrow$ )	(PSNR $\uparrow$ SSIM $\uparrow$ )	(PSNR $\uparrow$ SSIM $\uparrow$ )	(PSNR $\uparrow$ SSIM $\uparrow$ )	(PSNR $\uparrow$ SSIM $\uparrow$ )
Nearest	-	26.25 0.7372	24.65 0.6529	25.03 0.6293	22.17 0.6154	23.45 0.7414
Bilinear	-	27.55 0.7884	25.42 0.6792	25.54 0.6460	22.69 0.6346	24.21 0.7666
Bicubic	-	28.42 0.8101	26.00 0.7023	25.96 0.6672	23.14 0.6574	24.91 0.7871
SR-LUT	1274	29.82 0.8478	27.01 0.7355	26.53 0.6953	24.02 0.6990	26.80 0.8380
SP-LUT	5500	30.01 0.8516	27.21 0.7427	26.67 0.7019	24.12 0.7058	27.00 0.8430

Table 1: Quantitative results of

\times 4

super-resolution on five benchmark datasets. The best results are highlighted in bold , while the second best are marked with underline. Size denotes the LUT storage size.

Figure 3: Qualitative comparison of super-resolution results on Set14 and B100 using different LUT-based models and our IQ-LUT method.

DPFI	Res	Set5	Set14	B100	Urban100	Manga109
		30.63	27.59	26.90	24.50	27.69
✓		31.04	27.89	27.06	24.75	28.44
✓	✓	31.20	27.99	27.13	24.91	28.74

Table 2: Effect of DPFI and Residual Learning (Res) Modules on PSNR Across Benchmark Datasets for

\times 4

SR. The input quantization bit of every IQ-Block is 4.

NUQD	Set5	Set14	B100	Urban100	Manga109
	31.12	27.91	27.09	24.82	28.52
✓	31.17	27.95	27.10	24.85	28.65

Table 3: Effect of NUQD on PSNR Across Benchmark Datasets for

\times 4

SR. the first IQ-Block uses 4-bit input. All subsequent blocks operate use 3-bit input.

Image Super-Resolution

Image super-resolution reconstructs high-resolution images from lower-resolution observations, balancing detail, fidelity, and computational cost.

See all 248 papers

QuantSR+: Pushing the Limit of Quantized Image Super-Resolution Networks
May 21, 2026Haotong Qin, Xudong Ma, Xianglong Liu +4Quantization-Aware Training Low-Bit Quantization
ETH Zurich, Switzerland · Beihang University, China · Shanghai Jiao Tong University, China
ShiftLUT: Spatial Shift Enhanced Look-Up Tables for Efficient Image Restoration
Mar 1, 2026Xiaolong Zeng, Yitong Yu, Shiyao Xiong +4Image Restoration Edge Inference
Tsinghua University, Beijing · Kuaishou Technology, Beijing
Efficient INT8 Single-Image Super-Resolution via Deployment-Aware Quantization and Teacher-Guided Training
Apr 22, 2026Pham Phuong Nam Nguyen, Nam Tien Le, Thi Kim Trang Vo +1Efficient Neural Network Inference Quantization-Aware Training
University of Information Technology · 3Vietnam National University, Ho Chi Minh City · 2Ho Chi Minh City University of Technology

May 21, 2026J/K move · Enter open · S save

QuantSR+: Pushing the Limit of Quantized Image Super-Resolution Networks

Haotong Qin, Xudong Ma, Xianglong Liu, Jie Luo+3

ETH Zurich, Switzerland · Beihang University, China · Shanghai Jiao Tong University, China

Quantization-Aware Training Low-Bit Quantization Image Super-Resolution

Open paper AskPDF arXiv

Low-bit quantization is widely used to compress super-resolution (SR) models and reduce storage and computation costs for deployment on resource-limited devices. However, when SR models are pushed to ultra-low precision (2-4 bits), performance can drop sharply due to diminished representational capacity and the detail-sensitive nature of SR. To address these issues, we propose QuantSR+, a unified framework that improves quantization operators, network design, and training optimization, achieving better trade-offs between accuracy and efficiency than prior low-bit SR methods. QuantSR+ mainly relies on three technical contributions: (1) Redistribution-driven Bit Determination (RBD), which reshapes quantization distributions in both forward and backward passes to preserve representation fidelity; (2) Quantized Slimmable Architecture (QSA), which begins with an over-parameterized model and progressively prunes less critical blocks to meet efficiency budgets while pushing the accuracy performance; and (3) Slimming-guided Function-localized Distillation (SFD), which enforces block-aware feature alignment via a direct loss and a progressive, function-local training schedule to capture quantization effects better and speed up convergence. Extensive experiments show that QuantSR+ achieves state-of-the-art performance against both specialized quantized SR methods and generic quantization approaches. For SwinIR-S on Urban100 (x4), it improves PSNR by 0.29 dB over the 2-bit SOTA baseline. Meanwhile, it delivers strong efficiency gains at 2-bit, reducing operations by up to 87.9% and storage by 89.4%. QuantSR+ is effective for both convolutional and transformer-based SR models, indicating broad applicability.

Abstract

Image Super-Resolution

Explore similar work

QuantSR+: Pushing the Limit of Quantized Image Super-Resolution Networks

ShiftLUT: Spatial Shift Enhanced Look-Up Tables for Efficient Image Restoration

Efficient INT8 Single-Image Super-Resolution via Deployment-Aware Quantization and Teacher-Guided Training