SwiftFusion: Towards Motion-Robust and Efficient 4K HDR Fusion

High Dynamic Range (HDR) imaging remains a fundamental challenge in camera system design. Although recent diffusion-based methods have achieved promising visual quality for HDR fusion, they still suffer from significant limitations in robustness to dynamic scenes and prohibitively slow inference speeds, which make them difficult to apply to high-resolution HDR fusion in practice. To address these issues, we propose SwiftFusion , an efficient generative framework designed for high-resolution HDR fusion in complex dynamic scenes. It decomposes 4K HDR fusion into a motion-aware low-resolution fusion module with video priors, and a high-frequency-guided decoding module for faithful 4K reconstruction. Specifically, first, we propose a Motion-Robust Fusion Module, which remodels dynamic HDR fusion as an exposure-conditioned pseudo-video generation problem. This module uses a tailored condition-injection design to adapt pretrained video model for robust large-motion deghosting. Second, we design a High-Frequency Guided Decoder that extracts structural details from 4K LDR inputs and incorporates them into the decoding process as guidance. With a lightweight multi-scale design, this module can efficiently reconstruct low-resolution latents to 4K HDR images. Third, to further accelerate inference, we develop an acceleration pipeline that introduces a Condition-Aware Sparse Attention mechanism to reduce attention overhead and distills the flow-matching backbone into a one-step generator for accelerated sampling. Extensive experiments demonstrate that SwiftFusion achieves excellent deghosting capability and structural fidelity in highly challenging dynamic scenarios. Notably, SwiftFusion generates a 4K HDR image in only 2.2 seconds , achieving an impressive 198 × inference speedup compared with existing generative HDR approaches. Code is available at https://openimaginglab.github.io/SwiftFusion/