Physical Exposure-Prior and Gradual-Patch Network for Burst HDR Reconstruction

Song Gao, Jiacong Tang, Tao Hu, Xiaowen Ma, Qingsen Yan; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2026, pp. 1841-1850

High dynamic range (HDR) reconstruction from multi-exposure bursts requires precise alignment and robust feature fusion to handle complex motion and varying illumination, particularly in the non-linear JPG domain. In prevailing HDR frameworks, substantial challenges arise from noise sensitivity and structural artifacts, where traditional training on fixed-scale patches often leads to convergence failure due to the vast domain gap across exposures. To address these issues, we propose an end-to-end reconstruction framework specifically optimized for the NTIRE 2026 Restore Any Image Model Challenge (Track 2). Our approach introduces a Luminance-Prior Guided Exposure Alignment module that explicitly models physical exposure priors to unify luminance levels, significantly reducing the search space for subsequent alignment. Furthermore, to enhance training stability while maintaining global structural integrity, we adopt a Curriculum-based Progressive Tiling Strategy that evolves the spatial receptive field throughout the learning process. The core architecture utilizes a hierarchical 4-level Self-Attentive UNet to efficiently aggregate long-range dependencies and restore high-frequency details. Extensive experimental results demonstrate that our method effectively suppresses ghosting artifacts and achieves superior restoration quality. These findings highlight the synergy between physical-based alignment and progressive training strategies in advancing HDR reconstruction for real-world dynamic scenes.