💻 computer science

NTIRE 2026 Challenge on Single Image Reflection Removal in the Wild: Datasets, Results, and Methods

This paper reviews the NTIRE 2026 challenge on single-image reflection removal in the wild, highlighting the introduction of the real-world OpenRR-5k dataset, the participation of over 100 teams, and the state-of-the-art performance achieved by the top-ranked methods.

Original authors: Jie Cai, Kangning Yang, Zhiyuan Li, Florin-Alexandru Vasluianu, Radu Timofte, Jinlong Li, Jinglin Shen, Zibo Meng, Junyan Cao, Lu Zhao, Pengwei Liu, Yuyi Zhang, Fengjun Guo, Jiagao Hu, Zepeng Wang, Fe

Published 2026-04-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jie Cai, Kangning Yang, Zhiyuan Li, Florin-Alexandru Vasluianu, Radu Timofte, Jinlong Li, Jinglin Shen, Zibo Meng, Junyan Cao, Lu Zhao, Pengwei Liu, Yuyi Zhang, Fengjun Guo, Jiagao Hu, Zepeng Wang, Fei Wang, Daiguo Zhou, Yi'ang Chen, Honghui Zhu, Mengru Yang, Yan Luo, Kui Jiang, Jin Guo, Jonghyuk Park, Jae-Young Sim, Wei Zhou, Hongyu Huang, Linfeng Li, Lindong Kong, Saiprasad Meesiyawar, Misbha Falak Khanpagadi, Nikhil Akalwadi, Ramesh Ashok Tabib, Uma Mudenagudi, Bilel Benjdira, Anas M. Ali, Wadii Boulila, Kosuke Shigematsu, Hiroto Shirono, Asuka Shin, Guoyi Xu, Yaoxin Jiang, Jiajia Liu, Yaokun Shi, Jiachen Tu, Shreeniketh Joshi, Jin-Hui Jiang, Yu-Fan Lin, Yu-Jou Hsiao, Chia-Ming Lee, Fu-En Yang, Yu-Chiang Frank Wang, Chih-Chung Hsu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to take a beautiful photo of a painting inside a museum. But there's a problem: you have to shoot through a thick glass case. The glass reflects the lights in the room and your own face, creating a messy "ghost" image that sits on top of the painting. Your goal is to magically erase those reflections and see the painting clearly again.

This is exactly what the NTIRE 2026 Challenge was all about. It was a high-stakes competition where the world's smartest computer scientists and engineers tried to teach computers how to do this "magic eraser" trick on a single photo.

Here is a breakdown of the paper in simple terms:

1. The Problem: The "Ghost" in the Machine

For years, computers have been getting better at removing reflections, but they mostly practiced on fake, computer-generated images. It's like a pilot training only in a flight simulator; they might be great at the controls, but they might crash when they hit real turbulence.

The organizers realized that real-world photos are messy. The reflections aren't just simple lines; they are complex, overlapping, and vary in brightness. They needed a way to test if these AI models could actually work in the real world, not just in a lab.

2. The New Playground: OpenRR-5k

To fix this, the organizers (led by experts from OPPO and universities) built a new, massive training ground called OpenRR-5k.

How they got the data: Instead of asking people to take photos of paintings and then manually remove the glass (which is impossible to do perfectly), they used a clever trick. They took photos of real scenes with reflections, then used a super-smart AI tool (built into OPPO phones) to "guess" what the clean image looked like. Then, human experts went in like digital art restorers, using tools like Photoshop to fix any mistakes and make the "clean" version perfect.
The Result: A library of 5,000 pairs of photos: one with the messy reflection, and one perfect "clean" version. This became the gold standard for the competition.

3. The Competition: The Great Reflection Race

Over 100 teams signed up, but only 11 made it to the final round. They were given the messy photos and had to use their AI models to produce the clean versions.

Think of it like a cooking competition. Everyone was given the same "dirty" ingredients (the reflection photos) and had to cook the best "clean" dish (the restored image). The judges didn't just use a ruler to measure the food; they actually tasted it (subjective evaluation) to see if it looked and felt natural.

4. The Winners and Their Secret Recipes

The top teams didn't just guess; they used sophisticated "recipes" (algorithms). Here are the winners and what made them special:

1st Place: RRay (The Two-Step Chef)
- The Strategy: They realized that trying to remove the reflection and fix the details in one go was too hard. So, they built a two-stage machine. First, a "rough draft" AI removes the big, obvious reflections. Then, a second, finer-tuned AI comes in to polish the details and fix any blurriness.
- Analogy: It's like sanding a piece of wood. You start with a coarse sandpaper to remove the big bumps, then switch to fine sandpaper to make it smooth.
2nd Place: Xreflect Master (The Giant Brain)
- The Strategy: They took a powerful existing model and made it "bigger" (using a larger neural network backbone). They also used a technique called "Diffusion Distillation."
- Analogy: Imagine a student (the AI) learning from a genius professor (a massive Diffusion model). The student doesn't just copy the answers; they learn the intuition of the professor to handle tricky, weird reflections that the student hasn't seen before.
3rd Place: AIIALab (The Multi-Stage Refiner)
- The Strategy: They used a three-step process. First, they learned the basics. Second, they used "adversarial" training (where the AI plays a game against a critic to make the image look more realistic). Third, they used a "depth score" to check if the image made physical sense.
- Analogy: It's like a writer who writes a draft, then hires an editor to fix the grammar, and finally hires a fact-checker to ensure the story makes sense.

5. The Results: What Did We Learn?

The competition showed that:

AI is getting scary good: The top teams produced images that were almost indistinguishable from the real thing.
Numbers aren't everything: Sometimes, a computer score (like PSNR) said an image was "good," but human judges thought it looked weird or blurry. This proves that for real-world use, we need to train AI to please human eyes, not just math equations.
The Gap is Closing: The methods used by the winners are now being used in real phones and apps to help people take better photos through windows.

The Bottom Line

This paper isn't just a list of math formulas; it's a story of how the AI community is moving from "theoretical perfection" to "real-world magic." By creating a tough, realistic test (OpenRR-5k) and bringing together the best minds, they pushed the technology forward. Now, when you take a photo of a shop window or a museum exhibit, your phone might soon be able to instantly erase the reflection and show you the true beauty behind the glass.

1. Problem Definition

Single Image Reflection Removal (SIRR) is a fundamental image restoration task aimed at recovering the clean transmission layer ( $T$ ) from an input image ( $I$ ) contaminated by reflections ( $R$ ) caused by transparent surfaces like glass.

The Challenge: The problem is inherently ill-posed because the input is a linear combination of the transmission and reflection layers ( $I = T + R$ ).
The Gap: While deep learning has advanced SIRR, most existing methods are trained on synthetic data or limited real-world datasets. They often fail to generalize to complex, uncontrolled "in-the-wild" scenarios due to a lack of high-quality, pixel-aligned real-world ground truth data.

2. Key Contributions

The paper presents three primary contributions:

OpenRR-5k Dataset: A novel, large-scale real-world dataset specifically designed for SIRR.
- Collection Protocol: Unlike traditional methods that rely on manual glass removal or black velvet cloth (which cause misalignment), the organizers used an AI-driven pipeline integrated into OPPO smartphones. This involved using YOLO for detection, LaMa for inpainting, and NAFNet for removal, followed by manual refinement in Photoshop/MeituPic to ensure high visual fidelity.
- Composition: 5,300 high-quality pairs total.
  - Training: 5,000 pairs (OpenRR-5ktrain).
  - Validation: 300 pairs (OpenRR-5kval).
  - Testing: 100 input-only images (OpenRR-5ktest) with no ground truth, reserved for blind evaluation.
- Diversity: Covers 70% landscapes, 18% objects, 8% humans, and 4% animals, with varied lighting (Daytime, Indoor, Nighttime).
The NTIRE 2026 Challenge: A competition that attracted over 100 registrations and 11 final teams, aiming to bridge the gap between academic research and industrial application.
Comprehensive Benchmarking: A detailed analysis of 11 state-of-the-art methods, establishing new baselines for real-world reflection removal.

3. Methodology & Challenge Results

The challenge evaluated methods using both objective metrics (PSNR, SSIM, LPIPS, DISTS, NIQE) on the validation set and subjective evaluation by five experts on the test set. The final ranking was determined by the subjective score ( $S = 0.25C + 0.25A + 0.5Q$ ), prioritizing perceptual quality.

Top Performing Teams and Methods:

1st Place: RRay (RRay)
- Method: RdNafNet, a two-stage cascaded network.
- Strategy: Uses RDNet as a backbone to remove prominent reflections in Stage 1, followed by NAFNet in Stage 2 to recover fine details lost in the first stage.
- Training: Combined OpenRR-5k with real-world (Real-Nature, RRW) and synthetic data. Used Test-Time Augmentation (TTA) and weighted fusion of adversarial and non-adversarial models.
- Result: Highest subjective score (4.45) and best PSNR (36.17).
2nd Place: Xreflect Master (Xiaomi)
- Method: RDNet-XL with Diffusion Distillation.
- Strategy: Scaled up the backbone to FocalNet-XL for better global context. Employed Knowledge Distillation from diffusion-based models (WindowSeat, DAI) to generate pseudo-ground truth for hard cases.
- Training: Progressive resolution training (384 $\to$ 512 $\to$ 768) and VAE-based domain alignment to mitigate artifacts.
3rd Place: AIIALab (Harbin Institute of Technology)
- Method: MS-RDNet (Multi-Stage Refinement).
- Strategy: Three-stage training: (1) Supervised pre-training on OpenRR-5k, (2) Perceptual fine-tuning with Adversarial Loss on Nature dataset, (3) Physical evaluation using depth-consistency scoring to merge models.
Other Notable Entries:
- VIP Lab: Used Complementary Mixture-of-Experts (CoME) and Complementary Cross-Attention (CoCA) in a dual-stream framework.
- YuFans: Built on RDNet with Frequency-Aware Fine-Tuning, introducing Focal Frequency Loss (FFL) to target specific frequency bands where residual reflections hide.
- PSU: Proposed DUSKAN, a dual-path U-Net using Kolmogorov-Arnold Networks (KAN) for adaptive activation functions and FFT for global frequency modulation.
- SiGMoid: A two-head Restormer predicting both transmission and reflection explicitly, utilizing "checkpoint soup" (averaging weights) for robustness.

4. Key Findings & Analysis

Correlation between Metrics: PSNR showed a strong correlation with subjective scores for top-performing methods, suggesting that pixel-level fidelity is crucial in real-world scenarios. However, PSNR alone was insufficient to distinguish between methods with similar scores but different perceptual qualities (e.g., over-smoothing vs. detail preservation).
Data Strategy: The most successful teams (RRay, Xreflect Master) leveraged external data (synthetic or other real-world datasets) and advanced training techniques (diffusion distillation, progressive resolution) rather than relying solely on the challenge dataset.
Post-Processing: Techniques like Test-Time Augmentation (TTA) (flips, rotations) and Weight Averaging (SWA, Model Soup) provided significant performance boosts (up to ~1.8 dB PSNR gain for YuFans).
Limitations: Lower-ranked methods often suffered from residual reflections, over-smoothing, or color distortions, highlighting the difficulty of handling diverse lighting and reflection intensities in the wild.

5. Significance

This paper marks a significant step forward in the field of image restoration by:

Providing a Real-World Benchmark: Moving beyond synthetic data to a high-quality, pixel-aligned real-world dataset (OpenRR-5k) that reflects actual photography challenges.
Validating Industrial Viability: The use of OPPO's mobile AI pipeline for ground-truth generation demonstrates a practical pathway for creating high-quality datasets in industry.
Advancing SOTA: The winning methods demonstrate that combining large-scale architectures, diffusion priors, and multi-stage refinement can achieve near-perfect reflection removal in complex environments.
Community Resource: The release of the dataset and the detailed analysis of 11 diverse approaches provide a solid foundation for future research in reflection removal and general image restoration.

The challenge concludes that while significant progress has been made, achieving consistent perceptual quality across all diverse real-world scenes remains an open challenge, requiring further innovation in model architecture and data synthesis.