Boosting Inverse Tone Mapping via Diffusion Regularization

Xin Lu, Yufeng Peng, Chengjie Ge, Zhijing Sun, Ziang Zhou, Zishun Liao, Zihao Li, Dong Li, Qiyu Kang, Xueyang Fu, Zheng-Jun Zha; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2025, pp. 5654-5663

Images captured by digital cameras are constrained in their ability to represent the full luminance range of real-world scenes, often resulting in saturated pixels. To solve this problem, this paper proposes "Boosting Inverse Tone Mapping method via Diffusion Regularization Learning". Specifically, we first introduce the diffusion generative training paradigm into the general restoration network to promote the reconstruction effect of HDR images. By designing a series of regularization schemes and adding them to the fine-tuning process, we can effectively maintain the generative prior of the image restoration model in the reconstruction task, thereby greatly improving the HDR image reconstruction quality of Inverse Tone Mapping. At the same time, we also propose a simple and efficient U-Net to achieve HDR reconstruction in resource-constrained platforms, which can maintain extremely high HDR reconstruction effects while ensuring lightweight and computationally efficient. Experimental results on official datasets confirm the superiority of our method.