NTIRE 2026 The 3rd Restore Any Image Model (RAIM) Challenge: Multi-Exposure Image Fusion in Dynamic Scenes (Track 2)

Introduction

• MEF and HDR combine differently exposed frames to recover highlight and shadow details beyond a single shot’s range [7, 23, 25].MEF typically operates in nonlinear sRGB for display-ready results, whereas HDR reconstruction works in the linear domain and preserves radiometric information for later processing [7].Source: §1 · p. 1
• Moving objects, camera shake, local lighting changes, and exposure differences cause misalignment, ghosting, color distortion, and structural artifacts in dynamic-scene MEF.Practical deployment also requires efficient, stable processing on high-resolution inputs.Source: §1 · p. 1
• Learning-based MEF methods including MEFLUT [20], RetinexMEF [3], and UltraFusion [7] improve detail preservation and robustness, while HDR methods use optical flow, CNNs, and Transformers to reduce motion artifacts [23] [25] [49].Many HDR solutions still rely on heavy alignment components.Source: §1 · p. 1
• The RAIM series addresses the gap between academic research and practical deployment, where diverse degradations, limited memory, and strict latency requirements constrain real camera pipelines.Previous RAIM editions targeted RGB- and RAW-domain restoration and attracted academic and industrial participation.Source: §1 · p. 1
• NTIRE 2026 extends the RAIM initiative to dynamic-scene MEF through a new benchmark focused on motion robustness, detail fidelity, artifact suppression, and efficiency.The report aims to establish a reproducible benchmark and encourage efficient, generalizable solutions for real-world mobile HDR imaging.Source: §1 · p. 2

2. Challenge Data

• The benchmark targets realistic dynamic capture, with moving objects, local appearance changes, and camera shake between bracketed exposures of the same scene.It evaluates fusion quality and robustness to misalignment and ghosting artifacts.Source: §2 · p. 2
• The training set contains 100 multi-exposure sequences with 7 exposure levels, paired with reference fusion images for supervised training and validation.The exposure diversity supports studying its effects and testing fusion-pipeline robustness under dynamic motion.Source: §2 · p. 2
• The test set contains 100 sequences with 5 exposure levels, constraining fusion relative to training and testing stability with less exposure information.Source: §2 · p. 2
• The data includes object motion, illumination variation, handheld camera jitter and global misalignment, plus local saturation and underexposure.These conditions make detail recovery, motion handling, and visual naturalness central challenges for practical HDR fusion.Source: §2 · p. 2

3. Evaluation Protocol

• PSNR, SSIM, and LPIPS against private reference fusion images are combined into the public leaderboard’s single ranking score.The protocol also includes organizer-side reproducibility checks.Source: §3 · p. 2
• DISTS, NIQE, subjective quality [15], runtime, parameter count, and FLOPs are recorded beyond the main ranking score.Source: §3 · p. 2
• Phase-2 submissions must include exactly 100 JPEGs named 001.jpg–100.jpg and a readme reporting runtime, test hardware, extra training data, and notes.The files must be submitted in one .zip archive.Source: §3 · p. 2
• Final ranking includes only participants who submit inference code and pretrained weights on time, pass organizer review, and reproduce leaderboard performance within a reasonable range locally.The final-phase package must include a runnable test.sh script.Source: §3 · pp. 2–3

4.1. WHU-VIP

• WHU-VIP adapts AFUNet [25] to three exposures and adds spatially adaptive modulation and reliability gating for dynamic-scene fusion.It uses the 0 EV reference and ±2 EV frames; the SFT module modulates features by local illumination, while DCM-RG predicts pixel-wise gates to suppress unreliable regions. A four-stage curriculum progresses from alignment and augmentation to perceptual enhancement and 8-bit quantization-aware training.Source: §4.1 · pp. 3–4

4.2. SHL

• SHL concatenates a downsampled full-resolution image as global guidance and uses Neighboring Pixel Relationships loss to preserve local structure and high-frequency detail.The team trains Restormer [61], Uformer [58], and MST++ [6] from scratch, then fuses their predictions by weighted averaging.Source: §4.2 · pp. 4–5, Figure 3 · p. 4

4.3. nunucccb

• The nunucccb team combines spatially aligned five-frame features with a CA-ViT reconstruction network to produce the HDR result.Spatial attention aligns non-reference features to reference feature f3; CA-ViT blocks model global dependencies, while dilated convolution expands the receptive field and residual connections preserve reference structure. Training uses L1 reconstruction and VGG19 perceptual losses.Source: §4.3 · p. 5

4.4. untrafusion

• untrafusion combines arbitrary-frame fusion with optical-flow alignment, luminance-prior exposure matching, and progressive tiling for reconstruction.Its Restormer-based symmetric encoder-decoder handles reconstruction, and patch sizes grow from 256^2 to 512^2 to 768^2 over three training phases to suppress artifacts.Source: §4.4 · p. 5

4.5. I2 Group & Transsion

• I2 Group & Transsion separates five exposures into wavelet low- and high-frequency branches, using a learned residual to refine the middle exposure’s high-frequency features.The low-frequency branch uses DehazeFormer-s; the high-frequency design aims to avoid ghosting from fusing all exposures while retaining detail absent when using only the middle exposure.Source: §4.5 · p. 6

4.6. NTR

• NTR fine-tunes a pretrained masked diffusion autoencoder as a time-conditioned U-Net and aligns its composite training loss with the challenge metric.The five-exposure, 15-channel model combines ℓ1, SSIM, and LPIPS losses; tiled inference with an 8-way geometric self-ensemble forms the final submission.Source: §4.6 · p. 6, Figure 7 · p. 6

• 58.889: WHU-VIP ranks first overall; SHL leads PSNR and SSIM on both stages, but WHU-VIP has lower LPIPS and the highest stage scores.The overall ranking reflects the balance across both evaluation stages under the official metric.Source: §5 · p. 6, Table 1 · p. 7
• Representative Test Stage 1 comparisons show stronger methods producing cleaner structures, fewer ghosting artifacts, and better local detail preservation in challenging dynamic regions.High-performing solutions must balance perceptual naturalness with structural fidelity.Source: Figure 8 · p. 7, §5 · p. 7
• The top three teams remain relatively stable across stages, while several lower-ranked submissions drop on Stage 2, underscoring difficulty with exposure variation, motion, and misalignment.Source: §5 · p. 7

7. Conclusion

• The challenge paired a dynamic-scene dataset with transparent evaluation and organizer-side reproducibility verification to identify accurate methods intended to be robust in real-world use.It targets practical HDR challenges including motion and misalignment.Source: §7 · p. 8
• Top submissions balanced fidelity and perceptual quality across varied conditions, while dynamic artifacts remain an ongoing challenge.The organizers hope the challenge and released resources will support reliable, high-quality fusion for dynamic computational photography.Source: §7 · p. 8
• Participating teams contributed methods and insights that the organizers hope will drive further research in dynamic-scene fusion.Source: §7 · p. 8