The joint research paper 'Gradient Knows Best: Mixed-Precision Quantization via Gradient-Guided Bit Allocation for Super-Resolution' by Jeon Ju-hyun, Park Yun-seo and Cho Sung-in has been accepted as a regular paper at CVPR 2026 (IEEE/CVF Conference on Computer Vision and Pattern Recognition), the most prestigious international conference in computer vision.

The paper addresses two problems that arise when image super-resolution (SR) models are compressed with post-training quantization (PTQ): (1) statistics-based estimation of layer sensitivity does not reflect actual quantization error well, so bit allocation is inefficient; and (2) because BatchNorm is commonly removed in SR, activation scales fluctuate from sample to sample, so a fixed quantization range causes substantial performance degradation.

To alleviate these, the team proposes a PTQ-based mixed-precision quantization (MPQ) framework whose core elements are (i) gradient-guided bit allocation (GBA), which allocates bits by directly estimating per-layer quantization sensitivity from the loss gradient with respect to bit-width; (ii) bit-aware fine-tuning, which fixes the bit-widths determined by GBA and then re-adjusts the quantization ranges of weights and activations through learning; and (iii) dynamic activation range normalization (DAN), which normalizes activations per sample and per channel before quantization and restores the original scale afterwards, reducing clipping and range imbalance.

Calibrating on DIV2K (100 images), the research showed improved PSNR and SSIM over existing techniques across a range of SR models (EDSR, RDN, SwinIR) and benchmarks (Set5, Set14, BSD100, Urban100, Manga109).

Figure 1. Overview of the methodology

According to the announcement of the CVPR 2026 programme committee, 16,092 papers went through the review process this year (excluding withdrawals and desk rejects), of which 4,090 were accepted, an acceptance rate of 25.42%. The poster / highlight / oral designations will be announced separately at a later date.

CVPR 2026 will be held in Denver, Colorado, United States, from Wednesday 3 June to Sunday 7 June 2026.

[References]

Paper title: Gradient Knows Best: Mixed-Precision Quantization via Gradient-Guided Bit Allocation for Super-Resolution

Authors: Kim Jun-young (joint first author, Dongguk University), Jeon Ju-hyun (joint first author, Sogang University), Park Yun-seo (joint first author, Sogang University), Ahn Sang-yeon (fourth author, Dongguk University), Oh Yong-seok (fifth author, Dongguk University), Kim Bo-kyung (sixth author, Dongguk University), Cho Sung-in (corresponding author, Sogang University)

IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2026 Website

Website : https://sites.google.com/view/csi2267svm/