UltraLED: Learning to See Everything in Ultra-High Dynamic Range Scenes

TL;DR

UltraLED tackles reconstructing ultra-high dynamic range scenes from a single short-exposure RAW image by decoupling exposure correction and denoising. It introduces a ratio map $S$ to perform local exposure balancing in the RAW domain, coupled with a brightness-aware noise model to guide denoising. A 9-stop synthetic data pipeline enables realistic training data, and a two-stage UNet-based reconstruction achieves superior fidelity across highlights and shadows compared with RGB and multi-frame methods. The work delivers a public dataset and code, demonstrating robust, motion-free UHDR reconstruction suitable for practical single-frame imaging in dynamic scenes.

Abstract

Ultra-high dynamic range (UHDR) scenes exhibit significant exposure disparities between bright and dark regions. Such conditions are commonly encountered in nighttime scenes with light sources. Even with standard exposure settings, a bimodal intensity distribution with boundary peaks often emerges, making it difficult to preserve both highlight and shadow details simultaneously. RGB-based bracketing methods can capture details at both ends using short-long exposure pairs, but are susceptible to misalignment and ghosting artifacts. We found that a short-exposure image already retains sufficient highlight detail. The main challenge of UHDR reconstruction lies in denoising and recovering information in dark regions. In comparison to the RGB images, RAW images, thanks to their higher bit depth and more predictable noise characteristics, offer greater potential for addressing this challenge. This raises a key question: can we learn to see everything in UHDR scenes using only a single short-exposure RAW image? In this study, we rely solely on a single short-exposure frame, which inherently avoids ghosting and motion blur, making it particularly robust in dynamic scenes. To achieve that, we introduce UltraLED, a two-stage framework that performs exposure correction via a ratio map to balance dynamic range, followed by a brightness-aware RAW denoiser to enhance detail recovery in dark regions. To support this setting, we design a 9-stop bracketing pipeline to synthesize realistic UHDR images and contribute a corresponding dataset based on diverse scenes, using only the shortest exposure as input for reconstruction. Extensive experiments show that UltraLED significantly outperforms existing single-frame approaches. Our code and dataset are made publicly available at https://srameo.github.io/projects/ultraled.

Figures (17)

• Figure 1: Visualization results of different methods for UHDR scene reconstruction.
• Figure 2: Overview of our framework, which is divided into two parts: 1) Training Data Synthesis Pipeline: clean and normally exposed image $I$ is used as input. The lighted image $I_{L}$ is generated by artificially amplifying the brightness in specific regions of $I$. The noise model is then applied to $I_{L}$ to synthesize the corresponding noisy and overexposed image $I_{NL}$. Meanwhile, $I_{L}$ is linearly scaled down by different factors and clipped to produce pseudo multi-exposure images from $EV$-$0$ to $EV$-$N$. These images are fused using exposure fusion mertens2007exposure to obtain $I_{LF}$, and a clean exposure-corrected map (ratio map $S$) is derived by dividing $I_{L}$ by $I_{LF}$; and 2) UHDR Reconstruction Pipeline: this pipeline consists of two stages, implemented using two UNet chen2018learning. First, the Ratio Map Estimator takes the noisy and overexposed RAW image $I_{NL}$ as input and outputs the ratio map $\widetilde{S}$. Then, $\widetilde{S}$ is used to correct the exposure of $I_{NL}$, producing a noisy but well-exposed image $I_{N}$, which is then passed into the RAW Denoiser for denoising guidance. Finally, the RAW Denoiser outputs a clean and well-exposed RAW image $I_{CF}$. Note that, in terms of inference, only the UHDR Reconstruction Pipeline is used.
• Figure 3: A thumbnail of our UHDR dataset (convert to RGB images for visualization), where the images in the top row are UHDR inputs and the images in the bottom row are ground truth. Our UHDR dataset comprises 24 UHDR scenes, the majority of which are captured under three lighting conditions (ratio $\times$50, $\times$100, and $\times$200). For each combination of lighting condition and ISO setting, we captured three RAW images of the same scene to ensure statistical reliability. In total, the dataset includes 585 paired RAW images for evaluation.
• Figure 4: Visualization results of different methods on the UHDR dataset. The methods labeled blue take a long-exposure noiseless image as input, the methods labeled red take a short-exposure RAW image as input, and the methods labeled green take a short-exposure image RGB as input. UltraLED achieves good visual effects in both bright and dark regions. Note that there may be slight differences in our tone because our training data underwent a reversed ISP process. However, this difference is generally negligible and aligns with real-world conditions.
• Figure 5: Visual comparison on the SID chen2018learning dataset. It is well-known that RGB-based methods zhang2019kindlingguo2020zerojiang2021enlightenganbai2024retinexmamba have limited ability to recover details in extremely low-light regions. Therefore, we only compare UltraLED with ELD wei2021physics that is a representative RAW-based approach. UltraLED not only performs effective denoising in the dark areas of the scene but also well preserves the details in bright regions, such as the textures of trees, windows, and wall surfaces under light.