Source-linked AI summary
NTIRE 2026 The 3rd Restore Any Image Model (RAIM) Challenge: Multi-Exposure Image Fusion in Dynamic Scenes (Track 2)
Lishen Qu, Yao Liu, Jie Liang, Hui Zeng, Wen Dai, Guanyi Qin, Ya-nan Guan, Shihao Zhou, Jufeng Yang, Lei Zhang, Radu Timofte, Xiyuan Yuan, Wanjie Sun, Shihang Li, Bo Zhang, Bin Chen, Jiannan Lin, Yuxu Chen, Qinquan Gao, Tong Tong, Song Gao, Jiacong Tang, Tao Hu, Xiaowen Ma, Qingsen Yan, Sunhan Xu, Juan Wang, Xinyu Sun, Lei Qi, He Xu, Jiachen Tu, Guoyi Xu, Yaoxin Jiang, Jiajia Liu, Yaokun Shi
TL;DR
Dynamic-scene multi-exposure fusion must combine bracketed exposures despite motion, illumination variation, camera shake, and resulting misalignment or ghosting. NTIRE 2026 introduces a benchmark and evaluation protocol for this setting, with methods assessed for fidelity, perceptual quality, efficiency, and reproducibility. The challenge results show that leading submissions balance fidelity and perceptual quality across varied conditions, while dynamic artifacts remain difficult.
Problem
Dynamic-scene MEF remains difficult because moving objects, camera shake, lighting changes, and exposure differences create misalignment, ghosting, and structural artifacts in practical imaging.
Method
The challenge introduces a dynamic-scene MEF benchmark with realistic data, composite PSNR/SSIM/LPIPS ranking, broader quality and efficiency measures, and reproducibility verification.
Results
Top submissions balance fidelity and perceptual quality across varied conditions, while stronger methods produce cleaner structures, fewer ghosting artifacts, and better local detail preservation.
Takeaways & Limitations
The released benchmark and resources support reproducible research on reliable, efficient fusion for dynamic computational photography and mobile HDR imaging.
Abstract
from arXiv · showhide
This paper presents NTIRE 2026, the 3rd Restore Any Image Model (RAIM) challenge on multi-exposure image fusion in dynamic scenes. We introduce a benchmark that targets a practical yet difficult HDR imaging setting, where exposure bracketing must be fused under scene motion, illumination variation, and handheld camera jitter. The challenge data contains 100 training sequences with 7 exposure levels and 100 test sequences with 5 exposure levels, reflecting real-world scenarios that frequently cause misalignment and ghosting artefacts. We evaluate submissions with a leaderboard score derived from PSNR, SSIM, and LPIPS, while also considering perceptual quality, efficiency, and reproducibility during the final review. This track attracted 114 participating teams and received 987 submissions. The winning methods significantly improved the ability to remove artifacts from multi-exposure fusion and recover fine details. The dataset and the code of each team can be found at the repository: https://github.com/qulishen/RAIM-HDR.
1. Introduction
The challenge targets dynamic-scene multi-exposure fusion, where motion, camera shake, lighting changes, and exposure differences create severe artifacts. NTIRE 2026 extends RAIM with a benchmark designed to narrow the gap between academic methods and practical mobile HDR deployment.
- Motivation: MEF and HDR reconstruction combine differently exposed frames to recover highlight and shadow details beyond a single shot.MEF typically operates in nonlinear sRGB, while HDR reconstruction preserves radiometric information in the linear domain.
- Motivation: Moving objects, camera shake, local lighting changes, and exposure differences cause misalignment, ghosting, color distortion, and structural artifacts.These conditions make dynamic-scene MEF harder than fusion in controlled static environments.
- Motivation: Real camera pipelines face diverse degradations, limited memory, and strict latency requirements, creating a gap between academic research and practical deployment.RAIM was launched to address real-world restoration under practical efficiency constraints.
- Challenge scope: NTIRE 2026 extends RAIM to dynamic-scene MEF through a benchmark emphasizing motion robustness, detail fidelity, artifact suppression, and efficiency.Track 2 is presented as highly relevant to mobile photography.
- Challenge scope: The report covers dataset design, evaluation, submissions, participating methods, and final results to establish a reproducible benchmark for dynamic-scene MEF.The stated goal is to encourage efficient and generalizable solutions for real-world mobile HDR imaging.
2. Challenge Data
Track 2 uses a dynamic multi-exposure dataset designed to reflect realistic capture, including moving objects, appearance changes, camera shake, misalignment, and ghosting. It provides 100 training sequences with seven exposures and 100 test sequences with five exposures.
- Dataset characteristics: Each sequence contains bracketed exposures of one scene captured under possible object motion, local appearance changes, and camera shake.The benchmark evaluates both fusion quality and robustness to misalignment and ghosting.
- Training set: The training set contains 100 multi-exposure sequences, each with 7 exposure levels.This supports studying exposure diversity and fusion robustness under dynamic motion.
- Training set: Training data includes reference fusion images for supervised training and perceptual-quality validation.
- Test set: The test benchmark contains 100 sequences, each with 5 exposure levels.Compared with training, fewer frames make fusion more constrained and test stability with less exposure information.
3. Evaluation Protocol
The challenge ranks submissions with a composite score based on PSNR, SSIM, and LPIPS against private reference images. Evaluation also considers perceptual quality, efficiency, and reproducibility for practical deployment.
- Ranking metrics: The public leaderboard combines PSNR, SSIM, and LPIPS computed against private reference fusion images.These objective metrics form the official quantitative ranking protocol.
- Additional evaluation: Beyond the ranking score, evaluation records DISTS, NIQE, subjective quality, runtime, parameter count, and FLOPs.These dimensions assess perceptual reliability and computational efficiency.
- Data examples: Figure 1 presents representative image sequences from varied scenes and times, emphasizing high-dynamic-range content.
- Reproducibility: Final ranking is restricted to submissions whose code passes organizer review and reproduces leaderboard performance locally within a reasonable range.
4. Submitted Methods
Submitted methods address dynamic-scene fusion through selective frame use, metric-aligned objectives, global guidance, attention, physical exposure priors, wavelet decomposition, and diffusion-based reconstruction. Their designs variously target misalignment, ghosting, detail preservation, efficiency, and variable exposure counts.
- WHU-VIP: WHU-VIP uses three frames—0 EV and ±2 EV—within AFUNet, adding SFT and reliability gating to adapt fusion to motion and exposure conditions.The gating mechanism predicts pixel-wise gates to suppress unreliable regions and prioritize robust features.
- WHU-VIP: WHU-VIP’s curriculum progresses from alignment to dynamic generalization, perceptual enhancement, and quantization-aware training.Augmentations simulate exposure perturbation, gamma changes, spatial misalignment, and channel shuffling.
- SHL: SHL concatenates a downsampled full-resolution image with the input and adds NPR loss to preserve local structural and high-frequency details.The strategy is designed to improve established restoration models with minimal computational overhead.
- nunucccb: nunucccb uses spatial attention to suppress misalignment from foreground motion, then reconstructs HDR outputs with CA-ViT blocks and dilated convolution.Residual connections preserve the structural integrity of the reference feature.
- untrafusion: untrafusion combines optical-flow compensation, hierarchical Restormer reconstruction, luminance-prior alignment, and progressively larger training patches.The framework accommodates the difference between seven training frames and five test frames.
- I2 Group & Transsion: I2 Group & Transsion separates low- and high-frequency information with Haar wavelets and refines high-frequency details through residual fusion.The method processes five exposures while using the middle exposure as the high-frequency reference.
- NTR: NTR concatenates five LDR exposures into a 15-channel tensor and predicts the fused RGB image with a time-conditioned encoder-decoder.Its design aligns training with the official challenge metric while correcting local motion artifacts and ghosting.
5. Results Overview
The verified results rank methods across two test stages, with WHU-VIP leading overall and stronger methods showing cleaner, more structurally faithful outputs.
- 58.889 final score places WHU-VIP first overall, balancing performance across both evaluation stages.SHL achieves the highest PSNR and SSIM, but WHU-VIP records lower LPIPS and higher stage scores.
- Visual comparisons in Test Stage 1 emphasize structural consistency, detail preservation, and ghosting suppression.
- Top-three rankings remain relatively stable across Test Stage 1 and Test Stage 2, unlike several lower-ranked submissions.This indicates that robustness across different test conditions influences leaderboard position beyond peak fidelity on one subset.
- Stronger methods produce cleaner structures, fewer ghosting artifacts, and better local detail preservation in challenging dynamic regions.The qualitative results also show that perceptual naturalness and structural fidelity must be preserved together.
6. Teams and Affiliations
The listed teams represent institutions across China and include university, laboratory, industry, and research affiliations.
- Affiliations include Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ) in Shenzhen, China.
- Fuzhou University and Imperial Vision Technology are listed as collaborating affiliations.
- Other affiliations span Northwestern Polytechnical University, Beijing Jiaotong University, the Chinese Academy of Sciences, Transition, and Hohai University.
7. Conclusion
The challenge combines a dynamic-scene dataset, transparent evaluation, and reproducibility verification to support practical multi-exposure fusion research.
- NTIRE 2026 RAIM Track 2 focuses on multi-exposure fusion in dynamic scenes with motion and misalignment challenges.
- The challenge collects methods that balance leaderboard accuracy with robustness in real-world use.
- Released challenge resources are intended to encourage reliable, high-quality fusion methods for dynamic computational photography.