In this paper, we present an overview of the NTIRE 2026 challenge on the 3rd Restore Any Image Model in the Wild, specifically focusing on Track 1: Professional Image Quality Assessment. Conventional Image Quality Assessment (IQA) typically relies on scalar scores. By compressing complex visual characteristics into a single number, these methods fundamentally struggle to distinguish subtle differences among uniformly high-quality images. Furthermore, they fail to articulate why one image is superior, lacking the reasoning capabilities required to provide guidance for vision tasks. To bridge this gap, recent advancements in Multimodal Large Language Models (MLLMs) offer a promising paradigm. Inspired by this potential, our challenge establishes a novel benchmark exploring the ability of MLLMs to mimic human expert cognition in evaluating high-quality image pairs. Participants were tasked with overcoming critical bottlenecks in professional scenarios, centering on two primary objectives: (1) Comparative Quality Selection: reliably identifying the visually superior image within a high-quality pair; and (2) Interpretative Reasoning: generating grounded, expert-level explanations that detail the rationale behind the selection. In total, the challenge attracted nearly 200 registrations and over 2,500 submissions. The top-performing methods significantly advanced the state of the art in professional IQA. The challenge dataset is available at https://github.com/narthchin/RAIM-PIQA, and the official homepage is accessible at https://www.codabench.org/competitions/12789/.
Figures & tables
Figure 1: An example of pairwise comparison under the MCQ pattern, with annotations on the rationales behind the selection of the portrait part of the curated dataset.
Figure 2: An example of pairwise comparison under the MCQ pattern, with annotations on the rationales behind the selection of the scenery part of the curated dataset.
Figure 3: The structured prompt utilized for the LLM-as-a-Judge protocol to evaluate semantic alignment and reasoning quality.
Rank
Team
Final Award Score
Phase 2 Score
Phase 3 Score
1
IH-VQA
0.7305
0.6943
0.7667
2
VCIP Pi Group
0.7253
0.6826
0.7679
3
I 2 Group
0.7040
0.6490
0.7590
4
fugui
0.6939
0.6640
0.7237
5
LZ
0.6823
0.6149
0.7497
6
ongaku
0.6641
0.6056
0.7227
Table 1: Award Ranking for the Grand Prize on RAIM Track 1 after organizer-side verification. The final ranking is determined from the scores on Phase 2 and Phase 3.
Rank
Team
Phase 3 Score
LLM Score
Accuracy
1
VCIP Pi Group
0.7679
0.5535
0.6733
2
IH-VQA
0.7667
0.4459
0.7129
3
I 2 Group
0.7590
0.4557
0.7030
4
LZ
0.7497
0.4008
0.7228
5
fugui
0.7237
0.4137
0.6832
6
ongaku
0.7227
0.4113
0.6832
Table 2: Leaderboard of the Phase 3, ranked by Phase 3 Score.
Rank
Team
Phase 2 Score
Accuracy
Conditioned NLG
1
IH-VQA
0.6943
0.9118
0.2050
2
VCIP Pi Group
0.6826
0.9314
0.1098
3
fugui
0.6640
0.8824
0.1751
4
I 2 Group
0.6490
0.8627
0.1743
5
LZ
0.6149
0.8235
0.1557
6
ongaku
0.6056
0.8039
0.1776
Table 3: Leaderboard of the Phase 2, ranked by Phase 2 Score.
Figure 4: Visual comparison of generated rationales and selections with ground-truth annotations and selections on the portrait part of the curated test set.
Figure 5: Visual comparison of generated rationales and selections with ground-truth annotations and selections on the scenery part of the curated test set.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6: Overview of the proposed dualbranch framework of IH-VQA for pairwise image quality assessment. The Answer Model (AM) performs robust preference prediction through contentaware branching and model fusion, while the Thinking Model (TM) generates expertstyle textual rationales by integrating prompt templates, qualityrelated features, and Answer Model outputs. The final output includes both the preferred image decision and its corresponding explanation.
Figure 7: Data construction pipeline of VCIP Group.
Figure 8: Agent framework of VCIP Group.
Figure 9: Training pipeline of VCIP Group
Figure 10: Overview of the proposed I 2 Group framework for NTIRE 2026 RAIM Track 1. Official merged detailed images are first split into paired single-view crops and reformatted into four training sets. A Qwen3-VL-8B-Instruct backbone is then progressively optimized through five stages (SFT -¿ GRPO -¿ GRPO Again -¿ SFT Refresh -¿ Final GRPO). Finally, three complementary expert checkpoints selected from Stages 1, 2, and 5 are ensembled by majority voting for robust prediction.
Stage
Training Type
Validation Accuracy
Selected as Expert
Stage 1
SFT (LoRA)
0.74 ( 75/102 )
Yes, MOE_3
Stage 2
GRPO
0.77 ( 79/102 )
Yes, MOE_1
Stage 3
GRPO Again
0.79 ( 81/102 )
No
Stage 4
SFT Refresh (LoRA)
0.81 ( 83/102 )
No
Stage 5
Final GRPO
0.84 ( 86/102 )
Yes, MOE_2
Ensemble
Majority Voting
0.86 ( 88/102 )
Final System
Appendix
Table 4: Stage-wise Validation Performance of I 2 Group
School of Computer Science, Northwestern Polytechnical University, Xi’an, China · The Hong Kong University of Science and Technology, Hong Kong SAR, China