Research Team Led by Professor Cho Sung-in of the Department of Artificial Intelligence Wins Paper Acceptance at IEEE/CVF CVPR 2026
A joint research paper by Professor Cho Sung-in’s research team (Master’s students Jeon Ju-hyun and Park Yun-seo) of the Department of Artificial Intelligence has been accepted as a regular paper for CVPR 2026 (IEEE/CVF Conference on Computer Vision and Pattern Recognition), the premier international conference in computer vision. The study, titled “Gradient Knows Best: Mixed-Precision Quantization via Gradient-Guided Bit Allocation for Super-Resolution,” introduces an innovative approach to enhancing image resolution through efficient neural network optimization.
This paper addresses two issues that arise when lightweighting image super-resolution (SR) models via Post-Training Quantization (PTQ): first, that statistics-based layer sensitivity estimation fails to accurately reflect actual quantization errors, leading to inefficient bit allocation; and second, that since Batch Normalization is commonly removed in SR, activation scales fluctuate across samples, resulting in significant performance degradation under fixed quantization ranges.
To mitigate these issues, the research team proposes a PTQ-based Mixed-Precision Quantization (MPQ) framework. Its core components include: first, Gradient-guided Bit Allocation (GBA), which directly estimates layer-specific quantization sensitivity using the loss gradient with respect to bit-width to allocate bits; second, Bit-aware fine-tuning, which fixes the bit-width determined by GBA and re-adjusts the quantization range of weights and activations through training; and third, Dynamic Activation Range Normalization (DAN), which normalizes activations per sample and channel before quantization and then restores them to their original scale to reduce clipping and range imbalance.
After calibration using DIV2K (100 images), this study demonstrated improvements in PSNR and SSIM compared to existing methods across various SR models (EDSR, RDN, SwinIR) and benchmarks (Set5, Set14, BSD100, Urban100, Manga109).
According to the CVPR 2026 Program Committee, 16,092 papers underwent the review process this year (excluding withdrawals and desk rejections), with 4,090 papers accepted for a 25.42% acceptance rate. Meanwhile, the classification into Poster, Highlights, and Oral sessions will be announced separately at a later date.
CVPR 2026 is scheduled to be held from June 3 to 7, 2026, in Denver, Colorado, USA.
Authors: Kim Jun-Young (Co-first author, Dongguk University), Jeon Ju-Hyun (Co-first author, Sogang University), Park Yoon-Seo (Co-first author, Sogang University), Ahn Sang-Yeon (4th author, Dongguk University), Oh Yong-Seok (5th author, Dongguk University), Kim Bo-Kyung (6th author, Dongguk University), Cho Sung-In (Corresponding author, Sogang University)
▶IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2026 Website: https://sites.google.com/view/csi2267svm/
[SEO 키워드]
CVPR 2026, Mixed-Precision Quantization, Image Super-Resolution
[Summary]
The research team won a paper acceptance at CVPR 2026 for developing an innovative Mixed-Precision Quantization framework that significantly improves Image Super-Resolution model efficiency.
This paper addresses two issues that arise when lightweighting image super-resolution (SR) models via Post-Training Quantization (PTQ): first, that statistics-based layer sensitivity estimation fails to accurately reflect actual quantization errors, leading to inefficient bit allocation; and second, that since Batch Normalization is commonly removed in SR, activation scales fluctuate across samples, resulting in significant performance degradation under fixed quantization ranges.
To mitigate these issues, the research team proposes a PTQ-based Mixed-Precision Quantization (MPQ) framework. Its core components include: first, Gradient-guided Bit Allocation (GBA), which directly estimates layer-specific quantization sensitivity using the loss gradient with respect to bit-width to allocate bits; second, Bit-aware fine-tuning, which fixes the bit-width determined by GBA and re-adjusts the quantization range of weights and activations through training; and third, Dynamic Activation Range Normalization (DAN), which normalizes activations per sample and channel before quantization and then restores them to their original scale to reduce clipping and range imbalance.
After calibration using DIV2K (100 images), this study demonstrated improvements in PSNR and SSIM compared to existing methods across various SR models (EDSR, RDN, SwinIR) and benchmarks (Set5, Set14, BSD100, Urban100, Manga109).
According to the CVPR 2026 Program Committee, 16,092 papers underwent the review process this year (excluding withdrawals and desk rejections), with 4,090 papers accepted for a 25.42% acceptance rate. Meanwhile, the classification into Poster, Highlights, and Oral sessions will be announced separately at a later date.
CVPR 2026 is scheduled to be held from June 3 to 7, 2026, in Denver, Colorado, USA.
Authors: Kim Jun-Young (Co-first author, Dongguk University), Jeon Ju-Hyun (Co-first author, Sogang University), Park Yoon-Seo (Co-first author, Sogang University), Ahn Sang-Yeon (4th author, Dongguk University), Oh Yong-Seok (5th author, Dongguk University), Kim Bo-Kyung (6th author, Dongguk University), Cho Sung-In (Corresponding author, Sogang University)
▶IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2026 Website: https://sites.google.com/view/csi2267svm/
[SEO 키워드]
CVPR 2026, Mixed-Precision Quantization, Image Super-Resolution
[Summary]
The research team won a paper acceptance at CVPR 2026 for developing an innovative Mixed-Precision Quantization framework that significantly improves Image Super-Resolution model efficiency.