Construction then Reconstruction: Content-Adaptive One-Step Feature Reconstruction for HDR Deghosting

HDR deghosting reconstructs a high dynamic range image from multiple LDR images captured at different exposures. However, scene motion, severe exposure variations, and local saturation often cause misalignment and missing content, making HDR reconstruction challenging. To address this, we propose Construction then Reconstruction for HDR Deghosting (CRHDR), which first constructs a target representation for high-quality HDR reconstruction and then reconstructs it from incomplete multi-exposure observations. Specifically, we use optical flow and valid region masks to selectively align and fuse the input exposures, producing a coarse HDR image, which is then paired with a high-quality HDR image to construct target features for reconstruction. Next, we introduce a content-adaptive refinement (CAR) module that selects different processing paths according to local content, allowing high-quality target features to be incorporated only where the coarse representation is insufficient. Finally, we perform one-step diffusion in feature space, directly mapping coarse features to high-quality features for HDR reconstruction. Extensive experiments on three public real-world benchmarks show that our method significantly outperforms existing state-of-the-art methods in both quantitative metrics and visual quality, while making inference up to 40× faster. Our code and models will be made publicly available.