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

• SIRR seeks to recover the transmission layer T from an input image I contaminated by reflections R from reflective surfaces.Source: §1 · p. 1
• Traditional SIRR methods use hand-crafted priors [77], limiting generalization, while deep learning-based methods model uncertainty in transmission estimation [76].Source: §1 · p. 1
• A new collection protocol addresses the labor and pixel-alignment challenges of real-world data, producing diverse paired datasets OpenRR-1k [75] and OpenRR-5k [4].OpenRR-1k [75] served as the benchmark dataset for the NTIRE 2025 SIRR challenge.Source: §1 · p. 1
• The challenge aims to evaluate models in real-world scenarios and narrow the gap between academic research and practical photography.Source: §1 · p. 1

2.1. Overview

• The challenge’s objectives are to provide a diverse real-world benchmark and objective protocols, spur architectures for complex reflection intensities, and connect laboratory research to consumer-grade imaging.Source: §2.1 · p. 3

2.2. OpenRR-5k Dataset

• OpenRR-5k ground-truth transmissions begin with OPPO reflection-editor outputs and receive professional manual refinement, addressing limitations of traditional glass-removal or black-cloth collection.The authors cite misalignment and color shifts from environmental factors as potential problems in conventional reflection-pair collection.Source: §2.2 · pp. 3–4
• OpenRR-5k provides 5,000 training pairs, 300 validation pairs, and 100 input-only blind-test images covering diverse, genuine real-world reflection scenarios.The 5,300 collected pairs were divided into 5,000 training and 300 validation pairs; the 100 test inputs were reserved for blind evaluation of method performance and generalization.Source: §2.2 · p. 4

2.3. Challenge Phases

• The challenge has development/validation and testing phases: participants can train on paired data, submit validation outputs for metric feedback, and provide test outputs, code, and fact sheets.The validation set provides blended inputs without references; final submissions include outputs for both validation and test sets.Source: §2.3 · pp. 4–5

2.4.1. Objective Metrics on Validation Set

• Validation outputs are compared with ground-truth references using PSNR, SSIM, LPIPS, DISTS, and NIQE.Source: §2.4.1 · p. 5

2.4.2. Subjective Evaluation on Test Set

• Five experienced practitioners visually assessed test-set results, and the final score is a weighted average of three evaluation criteria.The listed assessment aspects include sharpness, detail preservation, exposure, and contrast.Source: §2.4.2 · p. 5

2.5. Final Challenge Ranking

• Although objective validation metrics and subjective test scores were collected for all eleven teams, final challenge ranking used only test-set subjective scores.Source: §2.5 · p. 5

3. Challenge Results and Analysis

• The challenge drew over 100 registrations and 1,000 submissions; eleven teams completed testing, with RRay winning first prize, Xreflect Master and AIIALab second, and VIP Lab, YuFans, and KLETech-CEVI third.Participating teams submitted model outputs, source code, and technical fact sheets.Source: §3 · p. 5
• Visual comparisons show top-ranked methods removing reflections while preserving details and colors, whereas lower-ranked methods often leave reflections, over-smooth, or distort colors.VIP Lab, YuFans, and KLETech-CEVI suppress intense reflections but struggle to faithfully recover background details.Source: Figure 5 · p. 6, Figure 6 · p. 7
• RRay leads with subjective score 4.45 and PSNR 36.1688, while Xreflect Master ranks second at subjective score 4.31 and records PSNR 36.0496 and lowest DISTS 0.0127.RRay also reports SSIM 0.9758, LPIPS 0.0235, and DISTS 0.0135; AIIALab ranks third and shows consistently competitive results.Source: Table 2 · p. 8, §3 · p. 8
• The official challenge ranking uses subjective scores on the test set, alongside objective metrics computed on OpenRR-5kval and subjective assessments on OpenRR-5ktest.The final evaluation covers all eleven participating teams.Source: §3 · p. 8, Table 2 · p. 8
• PSNR moderately correlates with subjective performance among top methods, but comparable PSNR can yield different subjective scores, revealing a perception–distortion trade-off.The paper notes YuFans and VIP Lab as examples of teams with comparable PSNR but noticeably different subjective scores, and says pixel-level fidelity alone is insufficient for high perceptual quality.Source: §3 · p. 8
• Top-performing approaches consistently use multi-stage or refinement-based designs, and leading methods typically train with external data beyond OpenRR-5ktrain.These observations point to progressive restoration and large-scale, diverse data as factors for handling complex reflection patterns and achieving state-of-the-art performance.Source: §3 · p. 8, Table 3 · p. 9

4. Challenge Methods

• Section 4 consolidates participating methods’ technical specifications and performance characteristics in Table 3, then describes the eleven competing methodologies.Source: §4 · p. 10

4.1. RRay – A Two-Stage Cascaded Network for Single Image Reflection Removal in the Wild

• RRay’s RdNafNet cascades RDNet [82] to remove visible reflections with NAFNet [9] to restore fine details lost in single-stage removal.The second stage refines the first stage’s intermediate output.Source: Figure 7 · p. 10, §4.1 · p. 10
• RRay augments training with 289 Real-Nature image pairs, 29,771 RRW real-world pairs, and dynamically synthesized examples from 42,736 images.The synthesized samples mimic reflection degradation patterns including ghosting and local reflections.Source: §4.1 · p. 10
• RRay trains first on 384×384 patches for 100,000 iterations, then on 768×768 patches for 50,000 more with adversarial loss, and uses flip-based test-time augmentation.The first phase uses MSE, gradient, and VGG perceptual losses; the submission also fuses outputs from adversarially trained and original models.Source: §4.1 · p. 10

4.2. Xreflect Master – Scaling RDNet with Diffusion Distillation for Reflection Removal

• Xreflect Master replaces RDNet’s FocalNet-L backbone with FocalNet-XL, strengthening multi-scale representation and global context while adding moderate computational overhead.The authors report consistent gains on validation and test sets.Source: Figure 8 · p. 10, §4.2 · p. 10
• Xreflect Master trains on OpenRR-5k alongside paired real-world data from RR4K (1,230 pairs), RRW [88] (3,000), PRR [81] (109), and DRR [25] (3,000).The RRW samples are a curated high-quality subset; the DRR samples are carefully selected and well-aligned.Source: §4.2 · p. 10
• Xreflect Master creates 1,000 hard cases using diffusion methods including WindowSeat [78] and DAI [25] to generate reflection-free pseudo-ground truth, with VAE reconstruction aligning inputs to supervision.The method targets extreme and rare reflection cases and seeks to distill diffusion models’ generative priors while mitigating VAE artifacts and domain gaps.Source: §4.2 · pp. 10–11
• Xreflect Master progressively trains at 384, 512, and 768 resolutions, then distills from a diffusion-based teacher for 20 epochs using hard cases mixed with original data.It uses flip-based test-time augmentation and averages predictions.Source: §4.2 · p. 11

4.3. AIIALab – MS-RDNet: A Multi-Stage Refinement RDNet with Adversarial Perception and Depth-Consistency Scoring

• AIIALab’s MS-RDNet combines supervised OpenRR-5k pretraining, Nature-dataset adversarial perceptual fine-tuning, depth-based scoring across phases, and model merging.The team also supplemented training with 200 additional real pairs from Nature [34].Source: §4.3 · p. 11

4.4. VIP Lab – Complementary Mixture-of-Experts and Complementary Cross-Attention for Single Image Reflection Separation in the Wild

• VIP Lab’s dual-stream separation framework combines Complementary Mixture-of-Experts and Complementary Cross-Attention, building on prior work.Figure 10 presents the overall architecture.Source: §4.4 · p. 11, Figure 10 · p. 11
• VIP Lab trained first on 7,643 synthetic and 290 real pairs, then fine-tuned on 4,950 OpenRR-5k pairs and validated on 50 pairs.Fine-tuning used 768×768 patches; test-time augmentation averaged predictions from the original image and its horizontal and vertical flips.Source: §4.4 · pp. 11–12

4.5. YuFans – Frequency-Aware Fine-Tuning with Post-Training Optimization for Reflection Removal

• YuFans fine-tuned an SIRS-pretrained RDNet on OpenRR-5k with spatial losses, then added Focal Frequency Loss to emphasize underperforming bands where residual reflections concentrate.Stage 1 reached 32.37 dB validation PSNR after 20 epochs; the architecture includes a FocalNet-Large backbone, RevCol body, three NAFBlock decoders, and a frozen ConvNext classifier.Source: §4.5 · p. 12, Figure 11 · p. 12
• Training only on OpenRR-5k yielded 32.77 validation PSNR, whereas adding synthetic SIRS data lowered it to 31.94, and MixUp also hurt performance.The authors attribute these degradations to domain gap and MixUp disrupting the physical blended-transmission relationship.Source: §4.5 · p. 12
• SWA added +0.05 dB over the best single checkpoint on Codabench, while 8× geometric TTA consistently boosted PSNR by ∼1.8 dB.The final submission used the same SWA checkpoint for validation and test.Source: §4.5 · p. 12
• SWA achieved 34.90 dB versus 34.88 dB for Model Soup, while multiscale TTA reduced performance by 2.48 dB.SWA averaged the top five epoch checkpoints selected by validation PSNR; multiscale TTA degraded quality through interpolation blur.Source: §4.5 · p. 12

4.6. KLETech-CEVI – XReflection-based Deep Network for Single Image Reflection Removal

• The decoder reconstructs the clean transmission layer through upsampling and convolutions, with skip connections preserving spatial information and improving gradient propagation.Source: §4.6 · p. 13
• The network estimates the reflection component, then subtracts it from the input to obtain the clean transmission image.This residual formulation is intended to simplify learning and focus the network on reflection artifacts.Source: §4.6 · p. 13
• Hierarchical features across multiple scales help the model capture both large reflection structures and subtle patterns.The encoder progressively downsamples while increasing feature channels, encoding context and reflection characteristics at different scales.Source: §4.6 · pp. 12–13
• The training objective is a weighted combination of losses, with λ1, λ2, and λ3 balancing their contributions.Source: §4.6 · p. 13

4.7. PSU – DUSKAN: Dual Spectral Kolmogorov-Arnold Network

• DUSKAN uses a symmetric four-level U-Net whose blocks blend a spectral-spatial path for global shifts with a KAN path for spatially varying degradations.The spectral path uses FFT magnitude modulation, positional encoding, and multi-scale depthwise convolutions; the KAN path learns polynomial-basis activations, and a learned per-stage gate blends the outputs.Source: §4.7 · p. 13, Figure 13 · p. 13
• DUSKAN trains from scratch using only official challenge training and validation data, then processes full-resolution images in one pass without overlapping patches or test-time augmentation.Training uses random 512×512 patches with rotations and horizontal and vertical flips.Source: §4.7 · p. 13

4.8. SiGMoid – Two-head Restormer with checkpoint soup and horizontal flip TTA

• SiGMoid’s four-stage Restormer predicts transmission and reflection separately, with L1, proxy-reflection, and consistency supervision.The reflection and consistency loss weights are λr = 0.2 and λc = 0.05; the authors report this decomposition loss was more stable than heavy perceptual or adversarial objectives.Source: §4.8 · p. 13
• SiGMoid uses only challenge training data, moderate augmentation, and official validation for model selection; continuation fine-tuning supplies nearby checkpoints for averaging.The main run lasts 200 epochs, followed by low-learning-rate continuation; stronger geometric or synthetic-domain perturbations tended to move samples away from official validation reflection patterns.Source: §4.8 · pp. 13–14
• The final SiGMoid submission averages three checkpoints with equal weights, a reliable practical gain over further continuation without averaging.The checkpoints are the best low-learning-rate fine-tuning checkpoint and continuation snapshots from epochs 40 and 48.Source: §4.8 · p. 14
• SiGMoid pads inputs to multiples of 8 and averages original and horizontal-flip predictions; D4 TTA and multi-scale inference did not improve validation scores.The input is restored without resizing and cropped back to its original size after padding.Source: §4.8 · p. 14

4.9. NTR – TimeDiffiT

• TimeDiffiT [61] follows MDAE pretraining with D2S-SFT fine-tuning on paired challenge images.The 4,750 paired training images contain blended and transmission layers; training uses t = 0, with no masking and σt = 0 during fine-tuning and inference.Source: §4.9 · p. 14
• MDAE pretraining [22] [61] combines spatial masking of 16×16 blocks with VE-SDE noise injection [55] and a reconstruction-plus-C2S loss.The masking ratio is sampled uniformly from 1% to 75%, enabling label-free pretraining on large-scale image corpora.Source: §4.9 · p. 14
• Fine-tuning uses AdamW with a 2 × 10−4 learning rate, cosine decay, batch size 8 per GPU, and random 512×512 crops.Training runs for 300 epochs, followed by 90 additional epochs incorporating the full training set.Source: §4.9 · p. 14
• The 142.5M-parameter TimeDiffiT [61] is a four-scale time-conditional U-Net with shifted-window TimeAttention blocks at each level and the bottleneck.A sinusoidal embedding and 2-layer MLP encode σt for AdaGN modulation; the full encoder-decoder is fine-tuned end-to-end.Source: §4.9 · pp. 14–15, Figure 15 · p. 14

4.10. refineX – Progressive Restormer U-Net with Multi-Loss Training for Single Image Reflection Removal

• SIRR-Net is a Restormer-style symmetric encoder-decoder U-Net for single-image reflection removal, operating at four resolution scales with pixel-unshuffle downsampling, pixel-shuffle upsampling, fused skip connections, and a global residual.Encoder stages E1–E3 use 64, 128, and 256 channels, followed by a 512-channel bottleneck.Source: §4.10 · p. 15
• Its stacked Transformer Blocks combine MDTA channel-wise attention, GDFN feature modulation, and SE channel recalibration; MDTA has O(C^2) rather than O((HW)^2) complexity.Depthwise convolutions capture local context, and gradient checkpointing is applied to the bottleneck during training.Source: §4.10 · p. 15
• refineX trains in three progressive stages at 128, 256, and 384 px, increasing resolution and perceptual and edge loss weights for the final fine-tuning stage.Stage 2 runs for 60 epochs and Stage 3 for 30 epochs; training uses DDP on three RTX 6000 Ada GPUs.Source: §4.10 · p. 15

4.11. ACVLAB – RDNet with Frozen DINOv2 Semantic Prior for Reflection Removal

• ACVLAB builds on RDNet [82], whose multi-column reversible encoder, FocalNet-L pretrained hierarchy extractor, and transmission-rate-aware prompt generator can yield blurry reconstructions under strong specular reflections.The authors attribute the blur to insufficient object-level semantics in FocalNet features.Source: §4.11 · pp. 15–16
• To reduce blurring, ACVLAB adds a frozen DINOv2 ViT-B/14 semantic branch, adapting its features at four encoder levels and fusing them by element-wise addition.The input is resized to 518×518; four lightweight adapters project features to 64, 128, 256, and 512 channels.Source: §4.11 · p. 16
• ACVLAB initializes from RDNet pretrained weights and fine-tunes the added adapters on OpenRR-1k and OpenRR-5k using MSE content, gradient-domain ℓ1, and VGG-19 perceptual losses.A two-phase strategy first loads the full RDNet [82] checkpoint, then fine-tunes with the DINOv2 branch to preserve encoder decoupling while steering the decoder toward sharper reconstructions.Source: §4.11 · p. 16

Comparison with Generative Methods

• The comparison evaluates generative baselines from the gpt-image and Nano Banana series alongside participant-proposed discriminative approaches.Source: §5 · p. 16
• Generative models can reconstruct transmission layers in complex reflection cases where competitive methods such as RRay and the OPPO-baseline struggle with overlapping high-frequency textures.The models leverage strong prior knowledge to infer the underlying layer.Source: §5 · p. 16
• Nano Banana models and gpt-image-2 restore sharp background details and suppress residual ghosting, while generative models preserve semantics in heavily occluded regions.Their semantic consistency may come with slight stochastic variations from the ground truth pixel-wise.Source: §5 · p. 16
• The zero-shot capability of gpt-image-1.5 suggests diverse image-text pretraining provides a prior for decoupling blended layers, but it can hallucinate misaligned textures and score lower on PSNR.The hallucinated textures may look plausible while deviating from ground-truth spatial structure.Source: §5 · p. 16
• Generative models may misidentify complex reflections as background objects, amplifying reflections or introducing ghosting artifacts absent from the input.Source: §5 · p. 16

• OpenRR-5k comprises 5,000 training pairs, 300 validation pairs, and 100 test images captured across diverse real-world environments.Source: §6 · p. 16
• The challenge drew participation from research and industry communities and demonstrated the effectiveness of recent deep learning approaches for reflection removal.Source: §6 · p. 16
• Consistent perceptual quality across diverse scenes remains a key challenge, motivating further work toward robust reflection removal for practical applications.The authors hope the dataset, benchmark, and analysis will encourage this research.Source: §6 · p. 16

Appendix B. Extended Visual Comparisons

• Figures 19–21 show visual comparisons on test0074, test0011, and test0015.Source: Figure 19 · p. 25, Figure 20 · p. 26, Figure 21 · p. 27