This paper presents a comprehensive review of the NTIRE 2026 3D Restoration and Reconstruction (3DRR) Challenge, detailing the proposed methods and results. The challenge seeks to identify robust reconstruction pipelines that are robust under real-world adverse conditions, specifically extreme low-light and smoke-degraded environments, as captured by our RealX3D benchmark. A total of 279 participants registered for the competition, of whom 33 teams submitted valid results. We thoroughly evaluate the submitted approaches against state-of-the-art baselines, revealing significant progress in 3D reconstruction under adverse conditions. Our analysis highlights shared design principles among top-performing methods and provides insights into effective strategies for handling 3D scene degradation.
Figures & tables
Figure 1: Visualizations of clean-degraded pairs in Track 1 & 2. Each track features 7 distinct scenes from the RealX3D benchmark [ 46 ] .
Track 1: 3D Lowlight Enhancement
Track 2: 3D Smoke Restoration
Method Name
Team Name
PSNR ↑
SSIM ↑
Method Name
Team Name
PSNR ↑
SSIM ↑
FuME-GS
DimV
23.3842
0.8019
GenSmoke-GS [ 4 ]
PLBBL
20.2061
0.7263
CISP-GS
DLMath_Vision
22.7783
0.7763
Smoke-GS [ 83 ]
XInsight Lab
18.6681
0.6909
TCIDNet-IBGS
Insta3D
21.6146
0.7122
Dehaze-then-Splat [ 9 ]
Hunan Duo
18.3816
0.6583
IDEAL
IDEAL
21.5019
0.7283
MSDG
AIIA_LAB
17.5486
0.6943
NAKA-GS [ 86 ]
BVG
21.4465
0.7786
DiT-IBGS
Insta3DDD
16.6687
0.6643
Table 1: Quantitative results of the 3DRR Challenge for Track 1 and Track 2, with teams presented in descending order of rank. Gray indicates teams that did not submit a valid factsheet and were therefore excluded from this report.
Figure 2: Overview of the proposed Fusion-Guided Multi-Stage Enhancement for 3DGS.
Figure 3: Overall architecture of the proposed multi-branch low-light 3DGS framework.
Figure 4: Overview of the proposed TCIDNet-IBGS.
Figure 5: Overall architecture of IDEAL. The Appearance MLP predicts Gain and Bias maps to absorb view-dependent artifacts from Pseudo RGB during training, supervised by Lcolor and Ldepth . At inference, the Appearance MLP is discarded.
Figure 6: Overview of the proposed NAKA-GS [ 86 ] pipeline, including NAKA-based enhancement, VGGT-based multi-view reconstruction, and Gaussian Splatting with PPM preprocessing.
Figure 7: Overview of GREP-GS. Pseudo-enhanced targets and relative depth priors are first extracted from low-light inputs, then 3DGS is optimized in a canonical display space using global and per-view adapters to handle view-specific enhancement bias.
Figure 8: Overview of IC-GS. Top: low-light enhancement via DarkIR with gamma boost, CIDNet-guided color correction, and SCUNet denoising. Bottom: multi-seed 2DGS training with ensemble rendering and gamma+std post-processing.
Figure 9: Overview of Space-GS. Low-light inputs are split into chrominance (HV) and intensity (I) streams by SAHT, enhanced through dual-stream processing with geometry-guided DMMA interaction, and reconstructed via inverse PSAHT.
Figure 10: The overall architecture of ELoG-GS [ 52 ] , featuring hybrid dual-branch reconstruction and post-enhancement stages.
Figure 11: Overview of AdaTone-GS. Adaptive pseudo-GTs supervise multi-feature Gaussians rasterized into RGB, depth, prior, noise, and illumination channels, then reconstructed via a cascaded denoiser and learnable tonemapper.
Figure 12: Overview of SLL-GS [ 26 ] . A three-stage low-light 3DGS pipeline: Stages 1–2 build geometry from COLMAP anchors with depth and reprojection priors; Stage 3 restores appearance via an illumination head and a YCbCr chroma residual.
Figure 14: Overview of DarkIR-GS. A DarkIR module restores illumination from low-light inputs, and differentiable rasterization back-propagates the RGB reconstruction loss into both the 3DGS parameters and the enhancement module for end-to-end training.
Figure 15: The overall framework of 3DLLR.
Figure 16: Overview of LLE-GS. An end-to-end framework that uses a transformer-based Retinex decomposition frontend to extract visual priors from low-light multi-view inputs, driving a view-adaptive Scaffolded 3DGS backend for high-quality NVS.
Figure 17: Overview of RunAI-Harmony3D.
Figure 18: Overall pipeline of the proposed AIC-GER method.
Figure 19: Overview of GenSmoke-GS [ 4 ] . Smoke-degraded multi-view images are sequentially restored via ConvIR-UDPNet, DCP-based dehazing, illumination correction, and MLLM-guided refinement under structural constraints, then reconstructed using 3DGS-MCMC with a FasterGS backend with ensembling.
Figure 20: Overview of Smoke-GS [ 83 ] . Hazy images are first enhanced by NanoBanana Pro, then a Smoke Medium Module encodes pixel ray directions via the Smoke-GS Encoder and predicts medium parameters through a Smoke Medium MLP. Finally, the predicted medium terms are fused with 3DGS rendering.
Figure 21: Overview of the MSDG pipeline. Multi-view hazy images are mildly pre-dehazed, fed into 3D-UIR for rendering, processed by a fine-tuned dehazing model, refined by Difix3D+, and finally fused and smoothed.
Figure 22: Overview of the proposed DiT-IBGS.
Figure 23: Overview of DePhy-GS. Smoke degradation is progressively removed via dehazing and deep enhancement, then reconstructed with a hybrid 3DGS representation and residual modeling, guided by a lightweight physical degradation model.
Figure 24: Overview of SmokeGS-R. Refined DCP inversion supervises a clean-only 3DGS with four donor branches as appearance priors; at inference, a geometric-mean reference from five renders harmonizes outputs in LAB space.
Figure 25: Overall architecture of SDG-GS. A clean 3D Gaussian scene representation is learned jointly with a simplified smoke forward model during training. At inference time, only the clean 3DGS renderer is used to produce the final restored test images.
Figure 26: Overall architecture of EE-GS. Degraded views are first restored by EE-Net, and then processed by Depth-AnythingV2 to estimate dense depth maps. Finally, the restored images together with the depth estimates are used in 3DGS for scene optimization.
Figure 27: Overview of 3DSmokeR. Stage 1 preprocesses smoke-degraded views with a fine-tuned MB-TaylorFormer; Stage 2 feeds the restored views into watersplatting [ 38 ] for 3D reconstruction.
Figure 28: Overview of MonoSmokeGS.
Figure 29: Overall pipeline of CPG-GS. Stage 1 utilizes a pre-trained Dehazeformer to remove scattering effects. Stage 2 explicitly leverages the restored images ( I^2D ) to extract reliable anchors via COLMAP, followed by per-scene Scaffold-GS optimization.
This paper presents our approach to the NTIRE 2026 3D Restoration and Reconstruction Challenge (Track 1), which focuses on reconstructing high-quality 3D representations from degraded multi-view inputs. The challenge involves recovering geometrically consistent and photorealistic 3D scenes in extreme low-light environments. To address this task, we propose Extreme Low-light Optimized Gaussian Splatting (ELoG-GS), a robust low-light 3D reconstruction pipeline that integrates learning-based point cloud initialization and luminance-guided color enhancement for stable and photorealistic Gaussian Splatting. Our method incorporates both geometry-aware initialization and photometric adaptation strategies to improve reconstruction fidelity under challenging conditions. Extensive experiments on the NTIRE Track 1 benchmark demonstrate that our approach significantly improves reconstruction quality over the baselines, achieving superior visual fidelity and geometric consistency. The proposed method provides a practical solution for robust 3D reconstruction in real-world degraded scenarios. In the final testing phase, our method achieved a PSNR of 18.6626 and an SSIM of 0.6855 on the official platform leaderboard. Code is available at https://github.com/lyh120/FSGS_EAPGS.
In this paper, we present a comprehensive overview of the NTIRE 2026 3rd Restore Any Image Model (RAIM) challenge, with a specific focus on Track 3: AI Flash Portrait. Despite significant advancements in deep learning for image restoration, existing models still encounter substantial challenges in real-world low-light portrait scenarios. Specifically, they struggle to achieve an optimal balance among noise suppression, detail preservation, and faithful illumination and color reproduction. To bridge this gap, this challenge aims to establish a novel benchmark for real-world low-light portrait restoration. We comprehensively evaluate the proposed algorithms utilizing a hybrid evaluation system that integrates objective quantitative metrics with rigorous subjective assessment protocols. For this competition, we provide a dataset containing 800 groups of real-captured low-light portrait data. Each group consists of a 1K-resolution low-light input image, a 1K ground truth (GT), and a 1K person mask. This challenge has garnered widespread attention from both academia and industry, attracting over 100 participating teams and receiving more than 3,000 valid submissions. This report details the motivation behind the challenge, the dataset construction process, the evaluation metrics, and the various phases of the competition. The released dataset and baseline code for this track are publicly available from the same GitHub repository, and the official challenge webpage is hosted on CodaBench.
This paper provides a review of the NTIRE 2026 challenge on real-world face restoration, highlighting the proposed solutions and the resulting outcomes. The challenge focuses on generating natural and realistic outputs while maintaining identity consistency. Its goal is to advance state-of-the-art solutions for perceptual quality and realism, without imposing constraints on computational resources or training data. Performance is evaluated using a weighted image quality assessment (IQA) score and employs the AdaFace model as an identity checker. The competition attracted 96 registrants, with 10 teams submitting valid models; ultimately, 9 teams achieved valid scores in the final ranking. This collaborative effort advances the performance of real-world face restoration while offering an in-depth overview of the latest trends in the field.