acceptodds
Withdrawn from ICLR 2027

Towards All-in-One Image Restoration with Unpaired Training

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

All-in-one (AIO) image restoration (IR) is crucial for addressing the diverse and complex degradations in real-world scenarios. However, existing AIO IR models are primarily trained with synthetic paired supervision and therefore suffer from severe domain limitations when applied to real-world data. Unpaired learning offers a promising alternative but introduces a significant challenge: removing degradations while faithfully preserving image content without pixel-level paired supervision. This challenge is further amplified in the AIO IR setting. To address it, we propose UADiff, an unpaired AIO IR framework that fine-tunes a pre-trained diffusion model into a one-step restoration backbone. Specifically, UADiff resolves two fundamental issues: degradation adaptation and unpaired constraint. For degradation adaptation, we design a Degradation-Aware Prompt Learning (DAPL) module leveraging SigLIP to learn a set of textual prompts representing degradation and clean concepts. These prompts identify degradations and guide a Degradation-Conditional Low-Rank Adaptation (LoRA) module, achieving degradation-adaptive restoration. For the unpaired constraint, we propose Content-Preserving Regularization (CPR), which promotes contrastive learning on degradation-invariant content features and guides the restored results away from degradation concepts. Extensive experiments demonstrate that UADiff achieves superior restoration performance and generalization, establishing a competitive baseline for unpaired AIO IR in real-world scenarios.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
All-in-One Image Restoration, Unpaired Learning, Diffusion Model

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_tJrdOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This article proposes UADiff, which is a non paired integrated image restoration framework based on a one-step diffusion model. It combines degradation aware fast learning, conditional LoRA adaptation, and content preserving regularization to handle multiple types of degradation without paired supervision.

Critical strengths

The ablation studies support the effectiveness of the proposed components. The paper is easy to follow.

Critical weaknesses
  1. In the current development context, exploring non-paired image restoration seems to have little significance. From the experimental results, it can be seen that the proposed method is far inferior to supervised methods, which makes the proposed method unconvincing.

  2. The proposed method is basically not innovative, and many ideas are borrowed from previous methods, such as DA-CLIP.

Questions for authors

Can the authors evaluate generalization to degradation types excluded from training?

Reviewer_ZeZLOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes UADiff, an unpaired all-in-one image restoration framework built upon a pretrained one-step diffusion model. Specifically, it introduces Degradation-Aware Prompt Learning (DAPL) and Degradation-Conditional Restoration (DCR) to enable adaptive restoration across different degradations. Additionally, Content-Preserving Regularization (CPR) is developed to balance degradation removal and content preservation under unpaired supervision.

Critical strengths
  1. Exploring unpaired learning for all-in-one image restoration is an interesting and meaningful research direction, as it reduces the reliance on paired training data and offers potential benefits for real-world applications.

  2. The proposed framework integrates degradation-aware prompt learning, conditional LoRA adaptation, and content-preserving regularization to address the challenges of unpaired image restoration. The overall design is reasonable and well-motivated.

Critical weaknesses
  1. The paper does not clearly explain the fundamental differences between unpaired all-in-one image restoration and conventional single-task unpaired restoration. In particular, what unique challenges arise when multiple restoration tasks are jointly learned without paired supervision? A more in-depth discussion of these challenges would better motivate the proposed framework.

  2. Several key techniques employed in this paper, such as contrastive learning, adversarial feature disentanglement, and degradation-aware guidance, have already been widely explored in previous single-task unpaired restoration methods. It remains unclear why these techniques are particularly effective in the all-in-one setting.

  3. The generalization capability to unseen degradation types remains unclear. The proposed DAPL module relies on a predefined set of degradation prompts learned from known degradation categories. Although the supplementary material includes experiments on mixed degradations, these combinations are still constructed from familiar degradation types. It would be valuable to investigate whether the framework can effectively handle previously unseen degradations or more complex real-world corruption distributions.

Questions for authors
  1. In Appendix D.2, Eq. (9) formulates the adversarial disentanglement objective as a min-max optimization, while the accompanying description states that the feature encoder aims to maximize the discriminator loss. Could the authors clarify the actual optimization directions for the feature encoder and discriminator, and how the adversarial gradients are implemented?

  2. In Appendix F, the authors state that positive samples for contrastive learning are selected from corresponding spatial locations in the target domain. However, Eq. (11) defines positive pairs between the degraded input and its restored output. Could the authors clarify how positive pairs are constructed in practice, particularly given that the degraded and clean training domains are explicitly unpaired?

Reviewer_vwKpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes UADiff, an unpaired all-in-one image restoration framework built on a pre-trained diffusion model. It aims to restore images affected by multiple types of degradation within a unified framework without relying on pixel-aligned training pairs. UADiff uses Stable Diffusion Turbo as its backbone and performs restoration in a single inference step.

UADiff consists of three main components. Degradation-Aware Prompt Learning (DAPL) uses SigLIP and learnable text prompts to extract degradation-related information. Degradation-Conditional Restoration (DCR) dynamically modulates the LoRA parameters of different network layers based on this information. Content-Preserving Regularization (CPR) combines adversarial feature disentanglement, contrastive learning, and semantic constraints to balance degradation removal and content preservation during unpaired training.

Experiments cover five restoration tasks: dehazing, deraining, denoising, low-light enhancement, and deblurring. UADiff outperforms the compared unpaired methods on standard benchmarks and achieves favorable perceptual quality on real-world datasets. Ablation studies and mixed-degradation experiments further support the effectiveness of its components.

Critical strengths
  1. Practically meaningful problem setting. The paper addresses unpaired all-in-one image restoration, aiming to reduce reliance on pixel-aligned training data. Extending unpaired learning to a unified framework that handles multiple degradation types is a worthwhile research direction with practical relevance.

  2. Coherent overall framework. UADiff integrates degradation-aware prompting and conditional restoration (DAPL and DCR) with content-preserving regularization (CPR) in a one-step diffusion restoration framework. The design addresses degradation adaptation and content preservation under unpaired supervision in a coherent manner.

  3. Competitive performance and extensive evaluation. UADiff achieves higher PSNR than all compared unpaired methods across five standard image restoration tasks. The paper also includes real-world evaluations, component-level ablations, and mixed-degradation experiments, providing substantial empirical support for the effectiveness of the framework.

Critical weaknesses
  1. The novelty of the core method is not clearly distinguished from prior work. The proposed DAPL–DCR degradation-adaptive restoration strategy shares a highly similar core idea with DOD (AAAI 2026, arXiv:2508.03373). Both use pre-trained models to perceive image degradation and degradation-conditioned modulation of LoRA for one-step all-in-one image restoration. However, the paper neither cites nor discusses this closely related prior work, nor does it clearly identify the independent novelty of DAPL–DCR relative to existing approaches. This weakens the novelty of the paper's second claimed contribution.

  2. High computational cost limits practical applicability. UADiff relies on a pre-trained diffusion model with approximately 1.273 billion parameters. For 512 × 512 images, it requires 2,816.42 GFLOPs and has an inference latency of 0.13 seconds. Both its computational cost and inference latency are slightly higher than those of CycleGAN-Turbo. This overhead limits its applicability to real-time processing and resource-constrained devices.

Questions for authors
  1. The paper presents the degradation-adaptive restoration strategy (DAPL–DCR) as one of its main contributions. However, DOD (AAAI 2026, arXiv:2508.03373) has already proposed a highly similar core idea: using pre-trained models to perceive image degradation and adaptively modulating LoRA based on degradation conditions for one-step all-in-one image restoration. This closely related prior work is neither cited nor discussed. Could the authors clarify the main novelty of their degradation-adaptive restoration strategy compared with DOD?

  2. UADiff improves restoration quality but still incurs substantial computational cost. How do the authors assess the trade-off between restoration quality and computational efficiency? From a practical perspective, which application scenarios are best suited to UADiff?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.