NTIRE 2026 3D Restoration and Reconstruction in Real-world Adverse Conditions: RealX3D Challenge Results
Shuhong Liu, Chenyu Bao, Ziteng Cui, Xuangeng Chu, Bin Ren, Lin Gu, Xiang Chen, Mingrui Li, Long Ma, Marcos V. Conde, Radu TimofteYun LiuRyo UmagamiTomohiro HashimotoZijian HuYuan GanTianhan XuYusuke KuroseTatsuya HaradaJunwei YuanGengjia ChangXining GeMache YouQida CaoZeliang LiXinyuan HuHongde GuChangyue ShiJiajun DingZhou YuJun YuSeungsang OhFei WangDonggun KimZhiliang WuSeho AhnXinye ZhengJiangxin DongKun LiYanyan WeiWeisi LinDizhe ZhangYuchao ChenMeixi Song Hanqing WangHaoran FengLu QiJiaao ShanYang GuJiacheng LiuShiyu LiuKui JiangJunjun JiangRunyu ZhuSixun DongQingxia YeZhiqiang ZhangZhihua XuZhiwei WangPhan The SonZhimiao ShiZixuan GuoXueming FuLixia HanChanghe LiuZhenyu ZhaoManabu TsukadaZheng ZhangZihan ZhaiTingting LiZiyang ZhengYuhao LiuDingju WangJeongbin YouYounghyuk KimIl-Youp KwakMingzhe LyuJunbo YangWenhan YangHongsen ZhangJinqiang CuiHong ZhangHaojie GuoHantang LiQiang ZhuBowen HeXiandong MengDebin ZhaoXiaopeng FanWei ZhouLinzhe JiangLinfeng LiLouzhe XuQi XuHang SongChenkun GuoWeizhi NieYufei LiXingan ZhanZhanqi ShiDufeng ZhangBoyuan TianJingshuo ZengGang HeYubao FuWeijie WangRunyi YangDeheng ZhangCunchuan Huang · 2026
Read NTIRE 2026 3D Restoration and Reconstruction in Real-world Adverse Conditions: RealX3D Challenge Results by Shuhong Liu, Chenyu Bao, Ziteng Cui, Xuangeng Chu, Bin Ren, Lin Gu, Xiang Chen, Mingrui Li, Long Ma, Marcos V. Conde, Radu TimofteYun LiuRyo UmagamiTomohiro HashimotoZijian HuYuan GanTianhan XuYusuke KuroseTatsuya HaradaJunwei YuanGengjia ChangXining GeMache YouQida CaoZeliang LiXinyuan HuHongde GuChangyue ShiJiajun DingZhou YuJun YuSeungsang OhFei WangDonggun KimZhiliang WuSeho AhnXinye ZhengJiangxin DongKun LiYanyan WeiWeisi LinDizhe ZhangYuchao ChenMeixi Song Hanqing WangHaoran FengLu QiJiaao ShanYang GuJiacheng LiuShiyu LiuKui JiangJunjun JiangRunyu ZhuSixun DongQingxia YeZhiqiang ZhangZhihua XuZhiwei WangPhan The SonZhimiao ShiZixuan GuoXueming FuLixia HanChanghe LiuZhenyu ZhaoManabu TsukadaZheng ZhangZihan ZhaiTingting LiZiyang ZhengYuhao LiuDingju WangJeongbin YouYounghyuk KimIl-Youp KwakMingzhe LyuJunbo YangWenhan YangHongsen ZhangJinqiang CuiHong ZhangHaojie GuoHantang LiQiang ZhuBowen HeXiandong MengDebin ZhaoXiaopeng FanWei ZhouLinzhe JiangLinfeng LiLouzhe XuQi XuHang SongChenkun GuoWeizhi NieYufei LiXingan ZhanZhanqi ShiDufeng ZhangBoyuan TianJingshuo ZengGang HeYubao FuWeijie WangRunyi YangDeheng ZhangCunchuan Huang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-04135.txt.
Introduction
Shuhong Liu and team reviewed the NTIRE 2026 3D Restoration and Reconstruction challenge, built around the RealX3D benchmark of low-light and smoke-degraded scenes. Of 279 registered participants, 33 teams submitted valid results, which the review evaluates against state-of-the-art baselines. Top methods showed significant progress in rebuilding 3D scenes under adverse real-world conditions. The paper draws out shared design principles among the leaders and lessons for handling degraded 3D scenes.
Catalog record
| CREATOR | Shuhong Liu, Chenyu Bao, Ziteng Cui, Xuangeng Chu, Bin Ren, Lin Gu, Xiang Chen, Mingrui Li, Long Ma, Marcos V. Conde, Radu TimofteYun LiuRyo UmagamiTomohiro HashimotoZijian HuYuan GanTianhan XuYusuke KuroseTatsuya HaradaJunwei YuanGengjia ChangXining GeMache YouQida CaoZeliang LiXinyuan HuHongde GuChangyue ShiJiajun DingZhou YuJun YuSeungsang OhFei WangDonggun KimZhiliang WuSeho AhnXinye ZhengJiangxin DongKun LiYanyan WeiWeisi LinDizhe ZhangYuchao ChenMeixi Song Hanqing WangHaoran FengLu QiJiaao ShanYang GuJiacheng LiuShiyu LiuKui JiangJunjun JiangRunyu ZhuSixun DongQingxia YeZhiqiang ZhangZhihua XuZhiwei WangPhan The SonZhimiao ShiZixuan GuoXueming FuLixia HanChanghe LiuZhenyu ZhaoManabu TsukadaZheng ZhangZihan ZhaiTingting LiZiyang ZhengYuhao LiuDingju WangJeongbin YouYounghyuk KimIl-Youp KwakMingzhe LyuJunbo YangWenhan YangHongsen ZhangJinqiang CuiHong ZhangHaojie GuoHantang LiQiang ZhuBowen HeXiandong MengDebin ZhaoXiaopeng FanWei ZhouLinzhe JiangLinfeng LiLouzhe XuQi XuHang SongChenkun GuoWeizhi NieYufei LiXingan ZhanZhanqi ShiDufeng ZhangBoyuan TianJingshuo ZengGang HeYubao FuWeijie WangRunyi YangDeheng ZhangCunchuan Huang |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 55037 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.04135
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
NTIRE 2026 3D Restoration and Reconstruction in Real-world Adverse Conditions: RealX3D Challenge Results Shuhong Liu, Chenyu Bao, Ziteng Cui, Xuangeng Chu, Bin Ren, Lin Gu, Xiang Chen, Mingrui Li, Long Ma, Marcos V. Conde, Radu TimofteYun LiuRyo UmagamiTomohiro HashimotoZijian HuYuan GanTianhan XuYusuke KuroseTatsuya HaradaJunwei YuanGengjia ChangXining GeMache YouQida CaoZeliang LiXinyuan HuHongde GuChangyue ShiJiajun DingZhou YuJun YuSeungsang OhFei WangDonggun KimZhiliang WuSeho AhnXinye ZhengJiangxin DongKun LiYanyan WeiWeisi LinDizhe ZhangYuchao ChenMeixi Song Hanqing WangHaoran FengLu QiJiaao ShanYang GuJiacheng LiuShiyu LiuKui JiangJunjun JiangRunyu ZhuSixun DongQingxia YeZhiqiang ZhangZhihua XuZhiwei WangPhan The SonZhimiao ShiZixuan GuoXueming FuLixia HanChanghe LiuZhenyu ZhaoManabu TsukadaZheng ZhangZihan ZhaiTingting LiZiyang ZhengYuhao LiuDingju WangJeongbin YouYounghyuk KimIl-Youp KwakMingzhe LyuJunbo YangWenhan YangHongsen ZhangJinqiang CuiHong ZhangHaojie GuoHantang LiQiang ZhuBowen HeXiandong MengDebin ZhaoXiaopeng FanWei ZhouLinzhe JiangLinfeng LiLouzhe XuQi XuHang SongChenkun GuoWeizhi NieYufei LiXingan ZhanZhanqi ShiDufeng ZhangBoyuan TianJingshuo ZengGang HeYubao FuWeijie WangRunyi YangDeheng ZhangCunchuan Huang Abstract 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. 1 Introduction Real-world visual data is inevitably captured under a wide range of adverse conditions that severely corrupt input image quality 6 ; 23 ; 5 ; 24 . As a fundamental task underpinning applications such as autonomous driving 79 ; 8 , robotic navigation 41 ; 48 , and augmented reality 40 ; 19 , 3D scene reconstruction demands high-fidelity multi-view inputs to produce accurate geometry and appearance. However, current reconstruction methods, i.e., Neural Radiance Field (NeRF) and 3D Gaussian Splatting (3DGS) 55 ; 32 ; 7 , are predominantly developed and benchmarked under controlled, clean-capture settings. Although recent studies 15 ; 35 ; 47 ; 38 ; 14 ; 49 ; 50 ; 16 have begun to address individual corruption types, achieving faithful reconstruction remains an open challenge 46 , and the scarcity of real-world degraded multi-view data further hinders progress. Figure 1 : Visualizations of clean-degraded pairs in Track 1 & 2. Each track features 7 distinct scenes from the RealX3D benchmark 46 . 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. Track 1: 3D Lowlight Enhancement Track 2: 3D Smoke Restoration Method Name Team Name PSNR \uparrow SSIM \uparrow Method Name Team Name PSNR \uparrow SSIM \uparrow 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 GREP-GS Wanderer 20.4584 0.7607 DEPHY-GS Diouj. El 15.3625 0.6574 IC-GS K7MQ 19.2141 0.6814 SmokeGS-R 22 Windrise 15.2174 0.6657 Space-GS XInsight Lab 18.8954 0.6793 HAD-GS ZZZ 14.8923 0.6103 ELoG-GS 52 LowLight Wizards 18.6626 0.6855 SDG-Smoke DLMath_Vision 14.4005 0.6341 AdaTone-GS SUSTech-PCL 17.7898 0.6762 EE-GS EE-GS 14.3759 0.6569 SLL-GS 26 3DV-Lowlight 17.7584 0.6382 3DSmokeR Helicopter 13.8623 0.5859 GammaGS HangFans 17.4242 0.6395 MonoSmokeGS MonoSmokeGS 13.6479 0.6251 DarkIR-GS FJNU-STAR 16.6617 0.6754 CPG-SGS AAA 13.5287 0.6181 3DLLR Stamina 16.5647 0.6914 DH-GS HNU 12.4948 0.5905 LLE-GS AAA 16.4346 0.6324 TD-GS FFFFYB 12.0852 0.5335 - DUTCS 16.2584 0.5765 Harmony3D RunAI 15.4799 0.5676 AIC-GER CC 13.9551 0.5924 To address the aforementioned challenges, we launched the first 3DRR Challenge at the 2026 NTIRE workshop. The goals of the challenge are threefold: (1) to highlight the critical problem of reconstruction under adverse conditions, (2) to offer a comprehensive evaluation benchmark, and (3) to drive research progress in this emerging field. The challenge features two tracks on two representative degradations, low-light and smoke. This challenge is one of the NTIRE 2026 Workshop associated challenges on: ambient lighting normalization 69 , raindrop removal 43 , smartphone ISP 58 , night-time image dehazing 1 , reflection removal 2 , shadow removal 70 , bitstream-corrupted 88 and UGC 42 video restoration, AI-generated image detection 27 , light-field 75 , efficient 62 , mobile 39 , infrared 45 , and 3D content 73 image super-resolution, efficient 77 and lowlight 10 image enhancement, image denoising 66 , efficient 21 and event-based 67 image deblurring, aberration correction 68 , deepfake detection 30 , photography retouching 18 , bokeh rendering 64 , video saliency prediction 56 , RAIM 59 , video interpolation 11 , cross-domain object detection 60 , X-AIGC quality assessment 51 , face restoration 72 and UGC images anomaly detection 84 . 2 Tracks and Competition Dataset The dataset used in the 3DRR Challenge is a subset of the RealX3D dataset 46 , which provides diverse real-world multiview image pairs of degraded and clean captures for faithful evaluation, as shown in Figure 1 . Tracks The challenge consists of two tracks: (1) 3D low-light enhancement and (2) 3D smoke restoration. Each track includes seven unique scenes from the RealX3D dataset 46 . For each scene, degraded views and ground-truth camera poses are provided as input. Participants are required to reconstruct the 3D scenes and submit rendered novel-view synthesis (NVS) results, which are evaluated against ground-truth (GT) clean captures. During the development phase, we fully released 1 validation scene with paired degraded and clean views for training purposes, along with 4 additional development scenes containing only degraded views. During the testing phase, we released 3 scenes with degraded views only. Each scene contains 3̃0 images for training and 5̃ images for NVS evaluation. Ranking criteria We evaluate the NVS results using PSNR and SSIM. To simplify the evaluation protocol, the final ranking is determined by the average PSNR across all scenes. SSIM is used only as a tiebreaker. Evaluation is conducted on the Codalab server. Participants are required to submit both their code and final results, which are checked by the organizers to ensure reproducibility. The challenge results of both tracks are summarized in Table 1 . 3 Track 1: 3D Lowlight Enhancement 3.1 FuME-GS: Fusion-Guided Multi-Stage Enhancement for 3DGS Mache You 1 , Zeliang Li 1 , Hongde Gu 1 , Jiangxin Dong 1 1 Nanjing University of Science and Technology Figure 2 : Overview of the proposed Fusion-Guided Multi-Stage Enhancement for 3DGS. Method The authors propose a multi-stage pipeline integrating multi-model enhancement, adaptive fusion, and 3DGS. As illustrated in Fig. 2 , input low-light images are first processed by multiple enhancement models (Retinexformer 3 , Zero-DCE 25 , ReDDiT 34 , and HVI-CIDNet 78 ) to generate a set of restored candidates. A region fusion strategy then combines these candidates to produce a high-quality enhanced image with improved illumination, reduced noise, and preserved structural details. For novel view synthesis (NVS), a 3DGS model is optimized using a fusion-guided point cloud initialization. Specifically, depth estimated from the low-light inputs provides reliable geometric structural cues, while the enhanced images