ExpandDiff improves recovery of hidden details in low dynamic range images
ExpandDiff: Dynamic Range Expanding Diffusion for Single-Image HDR Reconstruction
Computer Vision and Pattern Recognition
Summary
Images taken by regular cameras often lose details in very dark or very bright areas. This paper presents ExpandDiff, a new method that helps computers guess and fill in these lost details using a smart guessing process called diffusion. It works by learning from many examples and better handles cases where shadows or highlights are missing. Tests show it recovers image quality more accurately than previous methods. This can help make pictures look more natural and detailed without needing special cameras.
What this means in practice
- •For photo editing software developers: Integrate ExpandDiff to enhance details lost in shadows and highlights in single photos for improved HDR effects.$Commercial implications: This paper enables development of photo editing tools that offer advanced single-image HDR enhancement, appealing to photographers and consumers.
- •For mobile camera app engineers: Use ExpandDiff to reconstruct missing information in images taken by smartphones under challenging exposure conditions.
Authors
Mehmet Emre andıran, Zhuoqian Yang, Liying Lu, Mathieu Salzmann, Sabine Süsstrunk
Abstract
Single-image HDR reconstruction requires inferring missing detail while preserving the visible content of an LDR image. Differences in sensor dynamic range and exposure cause LDR images to lose varying amounts of information in shadows and highlights. We present ExpandDiff, a conditional diffusion pipeline that jointly reconstructs clipped shadows and highlights. To account for this variation, we introduce Dynamic Clipping Synthesis (DCS), which randomly samples shadow and highlight clipping percentiles when constructing training inputs from HDR targets. A pixel-space diffusion model guided by spatially-adaptive normalization then predicts perceptually encoded HDR through a bounded output head, reconstructing both clipping directions in one sampling trajectory. On the SI-HDR benchmark, ExpandDiff variants improve HDR reconstruction accuracy by 3.43 dB in PU21-PSNR over the strongest evaluated competing method, and by 7.34 dB under two-sided clipping. The code and supplementary material are available at https://memreandiran.github.io/expanddiff/.