NTIRE 2025 Challenge on Efficient Burst HDR and Restoration: Datasets, Methods, and Results

TL;DR

This paper introduces the NTIRE 2025 Efficient Burst HDR and Restoration Challenge, which targets efficient fusion of nine misaligned RAW frames into HDR RGB images under strict on-device constraints. A novel synthetic RAW HDR fusion dataset (300 training scenes, 20 validation/test scenes at 768×1536) with diverse degradations drives the competition, where teams must balance accuracy (PSNR/SSIM) and computational efficiency (≤30M parameters, ≤4T FLOPs). The six submitting teams reveal a shared two-stage paradigm (alignment followed by restoration) with varied architectures such as flow-guided deformable alignment, ConvNeXt-inspired blocks, and FFT-based transformers, achieving state-of-the-art results like 43.22 dB PSNR (SSIM 0.992) on the test set. The study provides practical insights for on-device burst HDR and restoration and points to future directions including runtime-aware metrics and richer ISP degradations to broaden evaluation beyond PSNR.

Abstract

This paper reviews the NTIRE 2025 Efficient Burst HDR and Restoration Challenge, which aims to advance efficient multi-frame high dynamic range (HDR) and restoration techniques. The challenge is based on a novel RAW multi-frame fusion dataset, comprising nine noisy and misaligned RAW frames with various exposure levels per scene. Participants were tasked with developing solutions capable of effectively fusing these frames while adhering to strict efficiency constraints: fewer than 30 million model parameters and a computational budget under 4.0 trillion FLOPs. A total of 217 participants registered, with six teams finally submitting valid solutions. The top-performing approach achieved a PSNR of 43.22 dB, showcasing the potential of novel methods in this domain. This paper provides a comprehensive overview of the challenge, compares the proposed solutions, and serves as a valuable reference for researchers and practitioners in efficient burst HDR and restoration.

Figures (8)

• Figure 1: Visualization of the dataset proposed in this competition. (a) Example of an HDR RGB GT image. (b) Cropped patch of (a). (c) Nine input RAW frames corresponding to (b), captured at three different exposure levels. Participants are required to fuse these noisy, misaligned and degraded frames to reconstruct the ground-truth RGB image in (b).
• Figure 2: The visualization of the participants' final results, along with the corresponding input reference frames and GT images. The first row displays the full input scenes, while the subsequent rows show cropped patches from the images. Each column represents the GT image, the input reference frame, and the results produced by the participants' models. Higher-performing models generate outputs that closely restore the GT, whereas the last two rows (e, f) illustrate cases where complete restoration was unsuccessful.
• Figure 3: The model proposed by the team ImvisionAI, named recursive multi-exposure alignment with spatiotemporal decoupling. They designed two stages of training the models, with each stage focusing on learning multi-frame alignment and image restoration, respectively. After the alignment module has been trained, it is frozen during the training of the restoration model in the second stage.
• Figure 4: The proposed model by DeepTrans, named flow-guided deformable alignment with channel-wise self-attention reconstruction.
• Figure 5: The model proposed by Team SimonLarsen. The full three-stage model architecture is shown in the top, the basic CN block on the right and the CNRDB block in the bottom. dw in the CN block denotes depth-wise convolution.