The Second LoViF 2026 Challenge on Real-World All-in-One Image Restoration: Methods and Results

arXiv:2607.21118 · cs.CV · Submitted 2026-07-23 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "The Second LoViF 2026 Challenge on Real-World All-in-One Image Restoration".

Jane: Here's the gist of the paper. This paper presents a review of the second LoViF Challenge on Real-World All-in-One Image Restoration, which aims to advance unified image restoration under diverse real-world degradation conditions, including blur, low-light, haze, rain, and snow.

Tom: The challenge provides a common benchmark for evaluating the restoration accuracy, robustness, and generalization capability of models across multiple degradation categories within a unified framework.

Jane: Challenge Objective and Benchmark The challenge investigates unified restoration under five common real-world degradation categories: blur, low-light, haze, rain, and snow.

Tom: The benchmark data is curated from the real-world paired data provided by FoundIR

27: and WeatherBench

18: , termed FoundIR-LoViF.

Title and authors: Tom: So the authors are setting up this challenge to see if a unified model can handle all those nasty things at once. It's not just one problem; it's five different degradation types in one go.

Jane: But the paper says that instead of building a separate model for each degradation type, we need one common framework that learns shared knowledge across those tasks twenty-one <ref:2607.21118#pg2>.

Lu: That means the goal is to see if a model can really generalize and adapt well to all these distinct characteristics while still using that shared knowledge effectively <ref:2607.21118#pg2>.

Meng: I'm worried about the real-world data gap, though. Synthetic data is usually clean, but real degradation has complex things going on in the camera pipeline or the scene itself <ref:2607.21118#pg3>.

Lalam: The paper acknowledges that synthetic datasets often differ substantially from what we see in reality because real degradations are affected by things like camera pipelines and illumination distributions <ref:2607.21118#pg2>.

Tom: So, the challenge isn't just about making a pretty picture; it’s about seeing if the model can actually handle the complexity of real-world degradation without needing a specific label for what kind of damage it is <ref:2607.21118#pg3>.

Jane: Exactly. It's testing whether these unified models can exploit shared knowledge across different tasks while still being adaptable to the unique features of each degradation type <ref:2607.21118#pg3>.

Lu: And they are using a balanced multi-task setting for their training, which is how they are trying to enforce that shared representation <ref:2607.21118#pg3>.

Meng: It sounds like a lot of training effort, but the real question is, can it actually perform when deployed on something that isn't perfectly balanced?

Lalam: The benchmark data they are using comes from FoundIR and WeatherBench, which covers five representative categories: blur, low-light, haze, rain, and snow <ref:2607.21118#pg3>.

The paper's summary: Tom: So the paper is summarizing this whole effort to develop unified restoration models for these five specific real-world degradation conditions. It's about finding that one model that works across all of them <ref:2607.21118#pg3>.

Jane: Essentially, they are contrasting this unified approach against the old way where you'd design and optimize a dedicated model for blur, then another for rain, and so on eight sixteen seventeen <ref:2607.21118#pg3>.

Lu: The unified setting evaluates whether a restoration model can effectively exploit shared knowledge across different tasks while maintaining sufficient adaptability to the distinct characteristics of each degradation <ref:2607.21118#pg3>.

Meng: And they are specifically looking at whether this common framework can actually work when it has to deal with real-world data that isn't perfectly clean <ref:2607.21118#pg3>.

Lalam: The authors state that the benchmark data are provided by FoundIR and WeatherBench, covering five representative categories of real-world image degradation, i.e., blur, low-light, haze, rain, and snow <ref:2607.21118#pg3>.

Tom: And they have a complete training set of twenty-four thousand paired low-quality and ground truth images for this evaluation <ref:2607.21118#pg3>.

Jane: So, the summary boils down to this: they want one model that can fix any of those five problems without needing extra task-specific training or labels <ref:2607.21118#pg3>.

Lu: That's right, it's about seeing if shared knowledge translates into good performance across all these different restoration tasks nine twenty-one <ref:2607.21118#pg2>.

Meng: If this unified model can generalize well to the real-world scenarios they are testing, that’s what matters for us when we think about deployment <ref:2607.21118#pg3>.

Lalam: The main focus is on developing a common framework to handle multiple degradation categories in one setting <ref:2607.21118#pg3>.

The paper's improvements: Tom: Now, the paper points out some specific ways they are trying to improve this unified restoration idea. They aren't just throwing a model at the data; they have these specific technical tweaks in mind.

Jane: What kind of improvements are they suggesting? Are we talking about better data handling or maybe smarter network design?

Lu: They suggest focusing on mechanisms that handle different degradation types specifically, like ReMamba emphasizing low-frequency information for haze and low-light restoration while keeping high-frequency structures for blur and rain <ref:2607.21118#pg3>.

Meng: That sounds like they are trying to be smart about which parts of the image they should focus on, rather than just applying one fix everywhere <ref:2607.21118#pg3>.

Lalam: LucidWorld introduces a wavelet-enhanced mixture of experts module that models global structures and illumination distributions under different degradation patterns <ref:2607.21118#pg3>.

Tom: And they also have something called the conditional NAFNet in NIT-Oita, which modulates normalized features in every restoration block using learned scale and bias parameters to show degradation-specific behavior <ref:2607.21118#pg3>.

Jane: That sounds like a way to give the network fine control over how it reacts based on what kind of damage it's seeing, even without an explicit category label <ref:2607.21118#pg3>.

Lu: And they are using dynamic degradation-aware optimization strategies where the loss weights are adjusted based on historical trends of validation losses <ref:2607.21118#pg3>.

Meng: Dynamic optimization is good for training, but how does that affect the inference side when we actually run this model in production? It adds complexity to the training, and then you have to make sure it's efficient enough <ref:2607.21118#pg3>.

Lalam: They are also incorporating frequency domain processing using Fourier Feature Convolution layers and FFT loss to boost frequency consistency and suppress artifacts, which is useful for rain streaks or snow particles <ref:2607.21118#pg3>.

Conclusion: Tom: So, wrapping this up, the paper is showing that you can have a unified model that tries to handle all five real-world degradations without needing task-specific checkpoints.

Jane: The key result they are highlighting is how these methods compete on the composite score, which includes PSNR and SSIM but also penalizes LPIPS <ref:2607.21118#pg3>.

Lu: The competition shows that while there are strong performers like Re:Pixel with a score of forty-two point two eight, there are still significant differences in the actual restoration accuracy and computational cost across all submissions <ref:2607.21118#pg3>.

Meng: From my side, I’m looking at the inference efficiency numbers. Re:Pixel takes about zero point one five seconds on an RTX four thousand ninety but DreamIR takes around three point eight six seconds per image Tom. That’s a big difference when you have a high volume of images to process <ref:2607.21118#pg3>.

Lalam: The paper concludes that effectively handling all five degradation types in one framework requires models to capture both the representations shared across tasks and the characteristics specific to each degradation type <ref:2607.21118#pg3>.

Tom: So, the overall implication is that we're seeing a lot of diversity in how these models balance restoration quality against speed and resource use <ref:2607.21118#pg3>.

Jane: It shows that while the unified framework is promising for generalization, we still need to figure out a better balance between restoration fidelity, perceptual quality, computational efficiency, and how well it handles those complex real-world conditions <ref:2607.21118#pg3>.

Lu: The challenge for future work seems to be achieving that better balance without needing explicit degradation labels or task-specific tuning nine twenty-one <ref:2607.21118#pg2>.

Meng: We need to see how these models perform when we move them out of the lab and into a real traffic situation with noisy data, not just perfect test sets <ref:2607.21118#pg3>.

Lalam: The Second LoViF two thousand twenty-six Challenge on Real-World All-in-One Image Restoration: Methods and Results is a major step in pushing for these more robust, generalizable AI restoration systems <ref:2607.21118#pg3>.

Nanjing University of Science and Technology

cs.CV

Submitted: 2026-07-23

Updated: 2026-11-13

Comments: ECCV 2026 Workshops; https://lowlevelcv.com/

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 67/100

The gist: This paper presents a review of the second LoViF Challenge on Real-World All-in-One Image Restoration…

Key concepts

FoundIR-LoViF Dataset
This is the specific benchmark data used in the challenge. It consists of 4,900 pairs of real-world low-quality images and their corresponding ground truths. This dataset ensures models are tested on diverse, realistic degradation conditions rather than just synthetic ones.
Composite Score
The evaluation metric used to judge model performance is a combination of three metrics: PSNR (Peak Signal-to-Noise Ratio) for pixel accuracy, SSIM (Structural Similarity Index) for structural quality, and LPIPS (Learned Perceptual Image Patch Similarity). This holistic score measures restoration fidelity and perceptual realism.
Mixture of Experts Framework
This approach involves a lightweight 'router' that analyzes the degraded input to predict which specific restoration expert—specialized for blur, haze, rain, etc.—should handle the image. This allows the model to specialize in different degradation types while still sharing general knowledge.
Frequency-Adaptive Modulation
This technique involves inserting modules at every feature scale that dynamically adjust how low-frequency (like haze) and high-frequency (like rain streaks) information is processed. It lets the same network adapt its internal processing to handle the unique frequency characteristics of different degradation types.

Terminology

Summary

This paper presents a review of the second LoViF Challenge on Real-World All-in-One Image Restoration, which aims to advance unified image restoration under diverse real-world degradation conditions, including blur, low-light, haze, rain, and snow. p. 0 The challenge provides a common benchmark for evaluating the restoration accuracy, robustness, and generalization capability of models across multiple degradation categories within a unified framework. p. 0

Challenge Objective and Benchmark

The challenge investigates unified restoration under five common real-world degradation categories: blur, low-light, haze, rain, and snow. The benchmark data is curated from the real-world paired data provided by FoundIR [27] and WeatherBench [18], termed FoundIR-LoViF. p. 1 The dataset contains 4,900 paired low-quality and ground-truth images with a resolution of 512 × 512 for training, resulting in a complete training set of 24,500 paired low-quality and ground-truth images. p. 2 Evaluation is performed using a composite score defined as Score= PSNR(Y) + 10 × SSIM(Y) -5× LPIPS. p. 2

Model Architectures and Approaches

Various teams propose different unified restoration frameworks, such as ReMamba, DreamIR, LucidWorld, Phoenix, SeeIR, MiVideo, MiAlgo LM, etc. Re:Pixel proposes ReMamba using a wavelet-driven state-space model. p. 3 REDnoteMediaLab suggests DreamIR as a pixel-level unified Diffusion Transformer that directly processes RGB pixels to avoid loss of high-frequency details. p. 5 LucidWorld utilizes a hierarchical encoder–decoder architecture with wavelet decomposition for progressive downsampling and feature decomposition. p. 6 Phoenix employs HyRoute, a two-stage mixture-of-experts framework utilizing degradation-aware hard routing and content-adaptive soft expert fusion. p. 8 SeeIR proposes RIRNet, combining a pretrained ConvNeXt backbone with a multi-scale encoder–decoder restoration branch. p. 9 MiVideo introduces AdaIGFormer, an adaptive illumination-guided Transformer that uses a task-aware degradation encoder to control five parallel processing branches. MiAlgo LM proposes a degradation-aware mixture-of-experts framework formulating restoration as a perceive-then-restore process. p. 7

Training Strategies and Optimization

Training strategies vary significantly across submissions, often employing multi-stage fine-tuning or progressive training schedules. For instance, DreamIR uses a five-stage progressive strategy involving flow-matching loss, LPIPS perceptual loss, and DINO perceptual supervision. p. 5 LucidWorld introduces a dynamic degradation-aware optimization strategy where loss weights are dynamically adjusted according to historical trends of validation losses. p. 7 MiVideo's AdaIGFormer involves six progressive stages, including training with RainRAG pairs and using FFT loss.

Performance and Trade-offs

The final results show intense competition among the leading solutions, with Re:Pixel achieving a final score of 42.28. The top three teams—Re:Pixel, REDnoteMediaLab, and LucidWorld—achieve scores of 42. p. 328, 41.81, and 41.40 respectively. p. 3 Despite strong performance, the submitted methods exhibit considerable differences in restoration accuracy and reported computational complexity. p. 3 The paper concludes that further improvements are needed to achieve a better balance among restoration fidelity, perceptual quality, computational efficiency, and generalization across complex real-world degradation conditions. p. 3

Key Findings on Model Design

Several methods focus on specific mechanisms to handle different degradation types; for example, ReMamba emphasizes low-frequency information for haze and low-light restoration while retaining high-frequency structures for blur, rain, and snow. LucidWorld introduces a wavelet-enhanced mixture-of-experts module to model global structures and illumination distributions under different degradation patterns. p. 7 The conditional NAFNet in NIT-Oita modulates normalized features in every restoration block through learned scale and bias parameters to exhibit degradation-specific restoration behavior.

Inference Efficiency

Inference efficiency varies, with ReMamba requiring approximately 0.15 seconds to process one image on an NVIDIA RTX 4090 GPU. p. 5 DreamIR achieves an average inference time of approximately 3. p. 686 seconds per image. p. 6 MiAlgo LM requires approximately 433 milliseconds to process one image of size 512 by 512 on a single NVIDIA RTX 3080 GPU. p. 5 The final model for O4A requires approximately 3 seconds per image on an NVIDIA A100 GPU with self-ensemble. p. 5

Final Conclusion on Complexity

The research highlights that effectively handling blur, low-light, haze, rain, and snow within a single framework requires models to capture both degradation-shared representations and degradation-specific characteristics. The competition reveals diverse trade-offs between model capacity, inference cost, and restoration quality. p. 3

Teams and Affiliations

The paper lists the teams involved in the challenge, including Re:Pixel from Hunan University Of Technology and Business. REDnoteMediaLab is affiliated with Xiaohongshu Inc. p. 3 LucidWorld members are affiliated with People’s Public Security University of China, Shanghai Police College. p. 3 Phoenix members are affiliated with Wuhan University. p. 7 SeeIR members are affiliated with Harbin Institute of Technology, Guangzhou University. p. 3 MiVideo is affiliated with Xiaomi Inc. p. 3 MiAlgo LM members are affiliated with Xiaomi Inc. p. 3 Pamy is affiliated with Sichuan University. p. 3 jason04 (Yu-Kai Wang) is from National Yang Ming Chiao Tung University. p. 0 NIT-Oita members are from National Institute of Technology. p. 3 ACVLAB members are affiliated with National Yang Ming Chiao Tung University, National Cheng Kung University. MoEFlow members are affiliated with Harbin Institute of Technology. p. 3 Zzz members are from Jilin University. p. 3 BaseLess members are from RIKEN BioResource Centre, Cochin University of Science and Technology. p. 3 O4A members are from Zhejiang University, The Hong Kong University of Science and Technology. p. 0 GKD IR members are from Wuhan University. p. 3 IKLab members are affiliated with IK Lab Inc. p. 3 sun is affiliated with Nanyang Technological University. p. 0

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Unified Model Development: The system can develop a unified model capable of recovering high-quality images from inputs affected by diverse real-world degradations, including blur, low-light, haze, rain, and snow. This is achieved by learning shared representations across multiple restoration tasks within a common framework.

  2. Robust Data Utilization: The system can utilize large-scale real-world paired data from datasets like FoundIR and WeatherBench to train foundation models for image restoration. This provides a substantial number of real-world samples with greater degradation diversity than synthetic degradations alone.

  3. Unified Evaluation Benchmark: The system can serve as a common benchmark for evaluating restoration accuracy, robustness, and generalization capability across multiple degradation categories within a unified framework. This is achieved by using the FoundIR-LoViF dataset and a consistent experimental setting for comparison.

  4. Adaptive Restoration Strategy: The system can employ adaptive degradation expert correction by using a feature router to predict pixel-level expert weights based on the input degradation prior, local high-frequency statistics, and restoration features. This allows the network to adapt its restoration behavior to different degradation types without explicit category labels during inference.

  5. Mixture of Experts Framework: The system can implement a mixture of experts framework where a lightweight degradation router first analyzes low-quality input and predicts whether it is affected by blur, low-light, haze, rain, or snow to dispatch the image to a dedicated restoration expert. This allows for task specialization while retaining cross-degradation knowledge sharing.

  6. Frequency Domain Consistency: The system can incorporate frequency domain processing using Fourier Feature Convolution layers and FFT loss to improve frequency consistency and suppress restoration artifacts. This is particularly useful for handling rain streaks, snow particles, blurred edges, and other unreliable high-frequency signals.

  7. Degradation-Aware Prompting: The system can utilize degradation-aware prompt interaction mechanisms where a prompt generator learns multiple degradation components to predict input dependent prompt weights from global feature representations. This allows the network to adapt its restoration behavior across different degradations at multiple network stages.

  8. Residual Correction with Semantic Verification: The system can perform spatially adaptive residual correction through one shared expert and several specialized experts, while a semantic feature verifier based on a frozen DINOv3 ConvNeXt backbone suppresses unsafe corrections. This prevents the residual correction from damaging image content by evaluating consistency with input structures and objects.

  9. Frequency-Adaptive Modulation: The system can insert frequency-adaptive modules at all feature scales to dynamically modulate low- and high-frequency information. This enables the same network to handle rain and snow patterns dominated by high-frequency components, haze and low-light degradations associated with low-frequency information, and blur affecting a wider frequency range.

  10. Blind All-in-One Restoration: The system can restore all five degradation categories using the same model parameters without requiring degradation labels, classifiers, or task specific checkpoints. This is achieved through a blind all-in-one restoration network that dynamically adapts its restoration features to the input degradation.

Sources

Related papers