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.
Figures & tables
Figure 1: Two representative data pairs from the 600 groups of Phase 1 training data. Each pair comprises a 1K-resolution low-light input image, a corresponding 1K-resolution flash portrait ground truth (GT), and a 1K-resolution person mask.
Figure 2: Two representative data pairs from the 100 groups of Phase 2 validation data. Similar to Phase 1, each group includes a 1K low-light input image, a GT, and a person mask. Compared to Phase 1, this split introduces more challenging scenarios, such as distant portraits, to improve the discriminative ability of the evaluation.
Figure 3: Two representative data pairs from the 100 groups of Phase 3 hidden test data. Each group comprises a 1K low-light input, a GT, and a mask, maintaining a scenario distribution and degradation difficulty consistent with the Phase 2 dataset.
Team
LPIPSperson↓
ΔEperson↓
GlobalScore ↑
Phase 2 Score
Rank
nunucccb
0.0266
7.19
0.7843
86.10
2
nanbei
-
-
-
85.21
3
SHL
0.0268
6.83
0.7416
84.91
4
kiritobryant
-
-
-
84.90
5
hezhaokun
0.0270
6.75
0.7388
84.88
6
KC110
0.0284
8.07
0.7647
84.33
7
Table 1: Phase 2 quantitative results on the validation dataset. The “-” symbol indicates that the team either failed to submit valid code during the reproducibility check or their reproduced score deviated significantly from the online leaderboard. ↑ indicates the higher the better, ↓ indicates the lower the better.
Figure 4: Visual comparison of the restoration results from the top 6 teams. The restored images are evaluated against the low-light input and the high-quality GT in terms of facial naturalness, detail preservation, lighting realism, and background cleanliness.
Team
Objective Score
UScore
Final Score
Rank
nunucccb
85.81
89.09
86.794
1
nanbei
84.66
90.00
86.262
2
KC110
85.37
87.27
85.940
3
NJUST-KMG
83.91
88.18
85.191
4
zte_cv
84.28
84.55
84.361
5
NTR
82.94
86.36
83.966
6
Table 2: Final comprehensive results of Phase 3. The Final Score is computed as the weighted sum of the reproduced Objective Score (70%) and the expert blind-test UScore (30%). ↑ indicates the higher the better, ↓ indicates the lower the better.