S2R-HDR: A Large-Scale Rendered Dataset for HDR Fusion

TL;DR

This work tackles the limited generalization of HDR fusion by introducing S2R-HDR, a 24,000-sample synthetic HDR dataset rendered with Unreal Engine 5 to capture diverse dynamic scenes. To bridge the synthetic-real gap, it proposes S2R-Adapter, a plug-and-play two-branch domain adaptation method that preserves shared knowledge while enabling domain-specific transfer, with test-time adaptation for unlabeled data. Empirically, models trained on S2R-HDR and augmented with S2R-Adapter achieve state-of-the-art HDR reconstruction on real datasets and demonstrate strong cross-domain generalization with minimal fine-tuning. The dataset and adaptation framework offer a practical path to robust HDR fusion in real-world, dynamic environments where data collection is challenging.

Abstract

The generalization of learning-based high dynamic range (HDR) fusion is often limited by the availability of training data, as collecting large-scale HDR images from dynamic scenes is both costly and technically challenging. To address these challenges, we propose S2R-HDR, the first large-scale high-quality synthetic dataset for HDR fusion, with 24,000 HDR samples. Using Unreal Engine 5, we design a diverse set of realistic HDR scenes that encompass various dynamic elements, motion types, high dynamic range scenes, and lighting. Additionally, we develop an efficient rendering pipeline to generate realistic HDR images. To further mitigate the domain gap between synthetic and real-world data, we introduce S2R-Adapter, a domain adaptation designed to bridge this gap and enhance the generalization ability of models. Experimental results on real-world datasets demonstrate that our approach achieves state-of-the-art HDR reconstruction performance. Dataset and code will be available at https://openimaginglab.github.io/S2R-HDR.

Figures (14)

• Figure 1: Comparing HDR fusion models kong2024safnet trained on our S2R-HDR dataset, with the proposed domain adapter S2R-Adapter, with the same model trained on previous SCT tel2023alignment and Challenge123 kong2024safnet datasets. Results show our dataset and training scheme can reduce ghosting artifacts under large motion (left) and recover very high dynamic range scenes, such as direct sunlight (right).
• Figure 2: The distribution of our S2R-HDR dataset and real captured HDR datasets Kalantari2017tel2023alignmentkong2024safnet. Following the approach outlined in shu2024towardsguo2023learninghu2022hdr, we first extract 7-dimensional features that capture key aspects of HDR, including the extent of dynamic range, intra-frame diversity, and the overall style of the HDR images. These features are then projected into a 2D space using t-SNE van2008visualizing for visualization.
• Figure 3: Illustration of our S2R-HDR dataset, covering both indoor and outdoor environments under diverse lighting conditions, including daytime, dusk, and nighttime, as well as various motion types such as humans, animals, and vehicles.
• Figure 4: Visualization of our sequence data and synthesized multi-exposure LDR images. Since the dataset consists of raw HDR sequences, it enables effortless data augmentation, such as brightness enhancement and motion amplitude adjustment.
• Figure 5: Structure of S2R-Adapter and t-SNE visualization of feature representations from different branches.