All-in-One Image Restoration (AIR)

• AIR is a unified image restoration framework that simultaneously addresses haze, rain, noise, blur, and low-light degradations.
• It leverages techniques like prompt-conditioning, latent inference, and model-based operators to adapt restoration across diverse tasks.
• Empirical results show enhanced PSNR and SSIM metrics alongside improved efficiency, though challenges remain for mixed degradation scenarios.

All-in-One Restoration (AIR), also termed All-in-One Image Restoration (AiOIR), denotes the construction of a single restoration model that handles multiple degradation types—such as haze, rain, snow, noise, blur, and low-light—without switching architectures or retraining per task. In its generic form, a degraded observation can be written as $$\mathbf{I}_{\deg} = \mathcal{D}_R(\mathbf{I}) + \epsilon$$, and AIR seeks a unified mapping $\hat{\mathbf{I}} = f_\theta(\mathbf{I}_{\deg})$ that operates blindly across multiple degradations and their composites. A related operator-centric formulation writes degradation as $I = H_t\{J\} + b_t$ and restoration as $J = G(I;\theta_t)$, emphasizing that the effective inverse should adapt across tasks while remaining within one framework (Sharif et al., 22 Sep 2025, Gao et al., 15 Jan 2026).

1. Definition, scope, and problem structure

AIR addresses a setting in which real images may suffer from unknown degradation type, mixed degradations, spatial heterogeneity, generalization demands, and efficiency constraints. The motivating use cases repeatedly cited across the literature include autonomous driving, surveillance, medical imaging, remote sensing, and consumer photography, where separate task-specific models are computationally impractical and break when degradations overlap (Sharif et al., 22 Sep 2025).

A central distinction between AIR and single-task restoration is that AIR must infer degradation characteristics from the degraded image itself. AutoDIR frames this as restoration under a general degradation operator $y = D(x; \theta) + n$, with a Blind Image Quality Assessment stage that infers degradation descriptions before restoration (Jiang et al., 2023). DAIR makes the blind setting explicit by requiring accurate restoration across multiple degradation types and their composites “without being told the degradation type” (Sharif et al., 22 Sep 2025). OPIR, by contrast, preserves the single-model objective but casts the problem through a physical degradation modeling perspective, in which the inverse operator is predicted in a task-aware manner (Gao et al., 15 Jan 2026).

The AIR literature therefore combines two persistent requirements. The first is unified representation: one model must encode common structure across degradations. The second is adaptive specialization: the same model must respond differently to haze, rain streaks, sensor noise, motion blur, low illumination, or their combinations. This tension between sharing and specialization underlies most architectural developments in the area.

2. Main architectural paradigms

Recent AIR systems can be organized into several recurring design paradigms. The categories below are directly reflected in the cited works and often coexist within hybrid systems.

Paradigm	Representative models	Core mechanism
Prompt- or instruction-driven conditioning	AutoDIR, PromptIR-style methods, DATPRL-IR	Learned prompts, text embeddings, or domain-aware task prompts
Hand-crafted or architectural priors	ADAIR-style frequency modulation, AnyIR, DPMambaIR	Frequency mining, state-space modulation, spatial-frequency adaptation
Latent inference and routing	DAIR, ClusIR, BaryIR	Latent priors, cluster-guided routing, barycenter-space decomposition
Operator- or model-based restoration	OPIR	Task-aware inverse degradation operator
Generative AIR	AutoDIR, UniLDiff, RestoreVAR, FAPE-IR	Latent diffusion, visual autoregression, MLLM-guided diffusion

Prompt-based AIR methods condition restoration with external text prompts, learned prompt tokens, or visual prompts. AutoDIR combines a CLIP-based Blind Image Quality Assessment stage with a text- and image-conditioned latent diffusion model and a Structural-Correction Module, thereby coupling open-set degradation detection with latent generative restoration (Jiang et al., 2023). DATPRL-IR extends prompt learning beyond a single image domain by introducing a task prompt pool, a domain prompt pool, and a Prompt Composition Mechanism that fuses task- and domain-level priors into a domain-aware task representation (Dong et al., 2 Mar 2026).

A second line of work seeks degradation awareness without explicit prompt dependence. DAIR criticizes both external text prompts and hand-crafted frequency heuristics as discrete assumptions that weaken generalization to unseen or mixed degradations, and instead reformulates AIR as learned latent prior inference (Sharif et al., 22 Sep 2025). AnyIR similarly rejects prompt modules and LLMs, using a joint embedding mechanism based on sub-latent splitting, gated degradation adaptation, and spatial-frequency fusion (Ren et al., 19 Apr 2025). DPMambaIR takes a state-space approach, using a pre-trained degradation extractor to modulate selective scan dynamics through a Degradation-Aware Prompt State Space Model (Liu et al., 24 Apr 2025).

A third family is explicitly model-based. OPIR predicts a task-aware inverse degradation operator implemented as a task-conditioned, multi-scale convolution kernel, and performs restoration by convolving this operator with the degraded image, followed by uncertainty-guided refinement in a second stage (Gao et al., 15 Jan 2026). This places AIR closer to classical inverse problems than to prompt-conditioned feature transformation.

Generative AIR has also diversified. UniLDiff uses Stable Diffusion XL with Degradation-Aware Feature Fusion and a Detail-Aware Expert Module to inject low-quality visual priors into latent diffusion (Cheng et al., 31 Jul 2025). RestoreVAR replaces iterative diffusion with visual autoregressive generation and reports over $10\times$ faster inference than latent-diffusion AIR baselines while retaining strong generalization (Rajagopalan et al., 23 May 2025). FAPE-IR adds a frozen MLLM planner that generates frequency-aware restoration plans and routes a LoRA-based Mixture-of-Experts inside a diffusion executor (Liu et al., 18 Nov 2025).

3. Degradation perception, routing, and representation learning

A defining theme in AIR is how the model represents degradation. DAIR formulates degradation as a latent variable $\mathbf{z}$ inferred by a VAE-like encoder $q_\phi(\mathbf{z}\mid \mathbf{I}_{\deg})$, with stage-wise latent codes $\{\mathbf{x}_\ell\}_{\ell=0}^3$ and a global bottleneck descriptor $\boldsymbol{\mu}$. These latent priors are then used in a structured “which–where–what” reasoning pipeline: which features to route, where to restore, and what content to reconstruct (Sharif et al., 22 Sep 2025). The resulting design combines latent-modulated encoder attention, learnable degradation maps $\hat{\mathbf{I}} = f_\theta(\mathbf{I}_{\deg})$0, and FiLM-like bottleneck modulation.

ClusIR replaces prompt or flat MoE routing with explicit degradation semantics via clustering. Its Probabilistic Cluster-Guided Routing Mechanism factorizes routing into a cluster posterior $\hat{\mathbf{I}} = f_\theta(\mathbf{I}_{\deg})$1 and a cluster-conditional expert posterior $\hat{\mathbf{I}} = f_\theta(\mathbf{I}_{\deg})$2, while the Degradation-Aware Frequency Modulation Module performs DWT-based decomposition, Fourier amplitude-phase fusion, and channel-wise high-frequency gating (Hu et al., 11 Dec 2025). This makes degradation recognition and expert activation separate operations rather than one shared softmax.

DPMambaIR injects degradation information into state-space dynamics. A pre-trained degradation extractor produces a 512-dimensional embedding $\hat{\mathbf{I}} = f_\theta(\mathbf{I}_{\deg})$3, which modulates the selective-scan parameters $\hat{\mathbf{I}} = f_\theta(\mathbf{I}_{\deg})$4, $\hat{\mathbf{I}} = f_\theta(\mathbf{I}_{\deg})$5, and $\hat{\mathbf{I}} = f_\theta(\mathbf{I}_{\deg})$6 in a Degradation-Aware Prompt State Space Model. A High-Frequency Enhancement Block then boosts $\hat{\mathbf{I}} = f_\theta(\mathbf{I}_{\deg})$7, where $\hat{\mathbf{I}} = f_\theta(\mathbf{I}_{\deg})$8 is global average pooling, to offset the high-frequency attenuation that arises in multi-task training (Liu et al., 24 Apr 2025).

AnyIR uses a different internal factorization. Its Degradation Adaptation Block skip-splits features into attention and gated paths, then applies a Gated Degradation Adaptation module whose temperature is adjusted from feature mean and standard deviation. Spatial fusion and FFT-based frequency fusion are then combined as $\hat{\mathbf{I}} = f_\theta(\mathbf{I}_{\deg})$9, giving the model explicit access to both spatial and spectral interactions (Ren et al., 19 Apr 2025).

Perceive-IR introduces quality perception as a distinct axis of representation. It trains CLIP prompts corresponding to “terrible,” “mediocre,” and “excellent” quality tiers and uses a CLIP-aware loss to move restored images toward the “excellent” prompt while pushing them away from the lower-quality prompts. In parallel, a Compact Feature Extraction module learns a degradation representation $I = H_t\{J\} + b_t$0 with a contrastive degradation-aware loss (Zhang et al., 2024). This suggests a shift from degradation type alone to joint modeling of degradation type and severity.

Other representation-learning approaches generalize this theme further. BaryIR decomposes the latent space into a continuous barycenter space $I = H_t\{J\} + b_t$1 for degradation-agnostic information and source-specific subspaces $I = H_t\{J\} + b_t$2 for degradation-specific semantics, using a multi-source latent optimal transport barycenter objective (Tang et al., 27 May 2025). UDAIR instead learns degradation pattern codebooks and aligns source and target embeddings by dynamically adapting target features toward source cluster structure using CORAL-based test-time adaptation (Fan et al., 28 May 2025). ClearAIR combines a DeQA-derived global quality embedding, SAM2-based semantic masks, DA-CLIP degradation prompts, and an internal clue reuse mechanism, organizing these as a “How–Where–What” hierarchy inspired by Human Visual Perception (Zhang et al., 6 Jan 2026).

4. Benchmarks, datasets, and evaluation protocols

AIR evaluation is unusually heterogeneous because the task spans multiple degradations, severity levels, and domain conditions. DAIR evaluates across six common restoration tasks, five compound settings, and unseen compound degradations, using PSNR and SSIM on paired sets, NIQE, MUSIQ, and LPIPS on real or unseen data, GFLOPs and inference speed for efficiency, and downstream object detection mAP with YOLOv12-L on a multi-degradation dataset (Sharif et al., 22 Sep 2025). OPIR adopts an all-in-one setting built from deraining, desnowing, and dehazing datasets, and also reports task-aligned results to show that a unified model can remain competitive in per-task evaluation (Gao et al., 15 Jan 2026).

WeatherBench addresses the data substrate itself rather than the architecture. It introduces a real-world all-in-one adverse weather benchmark with 42,002 paired degraded/clean images, standardized at $I = H_t\{J\} + b_t$3, spanning rain, snow, and haze under daytime and nighttime conditions. The train/test split is 41,402/600, and the acquisition system uses a fixed camera setup and a unified protocol for weather generation and ground-truth capture (Guan et al., 15 Sep 2025). Because many earlier AIR studies mix synthetic single-weather datasets with different resolutions, styles, and synthetic pipelines, WeatherBench isolates a major confound: cross-dataset domain gap.

This benchmark-centric perspective is important because AIR performance can depend not only on architectural design but also on whether the training set is a coherent multi-degradation domain. WeatherBench shows that methods trained on a unified real dataset often produce better real-world PSNR and comparable or slightly improved SSIM than the same methods trained on hybrid synthetic mixtures (Guan et al., 15 Sep 2025). A plausible implication is that future AIR comparisons will increasingly distinguish architecture effects from dataset-construction effects.

5. Empirical trends: performance, efficiency, and generalization

A recurring empirical result is that unified models can now match or exceed specialized systems on average metrics while also improving robustness to unseen conditions. DAIR reports 28.15 dB PSNR and 0.8829 SSIM averaged over six tasks, with an average PSNR improvement of 1.68 dB and 0.0598 SSIM over the best prior AIR method, using 18.08M parameters and 45.65 GFLOPs; it also reports 22.61 dB and 0.8277 SSIM on five compound degradations, 20.97 dB and 0.8415 SSIM on unseen compound degradations, and 34.9% AP at IoU $I = H_t\{J\} + b_t$4 for downstream detection on restored images (Sharif et al., 22 Sep 2025).

OPIR reports an average PSNR of 34.02 in its all-in-one derain/desnow/dehaze setting, outperforming PromptIR, VLU-Net, and Perceive-IR while running in 0.174 s compared with 0.682–1.012 s for the listed AIR baselines (Gao et al., 15 Jan 2026). AnyIR emphasizes efficiency even more directly, reporting around 82% fewer parameters and 85% fewer FLOPs while achieving SOTA performance in the all-in-one setting (Ren et al., 19 Apr 2025). RestoreVAR positions itself on the generative side of AIR and reports 0.201 s inference, 1.05 TFLOPs, and over $I = H_t\{J\} + b_t$5 faster inference than LDM-based AIR baselines such as Diff-Plugin, AutoDIR, and PixWizard, while remaining the strongest generative AiOR model on the listed benchmarks (Rajagopalan et al., 23 May 2025).

Generalization results are increasingly central. BaryIR shows strong performance on unseen real-world degradations such as O-HAZE and SPANet, on mixed degradations such as haze+rain and blur+noise, and on unseen noise levels $I = H_t\{J\} + b_t$6, which supports the claim that degradation-agnostic barycenter representations can improve out-of-distribution behavior (Tang et al., 27 May 2025). UDAIR explicitly reframes AIR under domain shift and uses codebook-based degradation pattern matching plus test-time adaptation to transfer from source datasets such as SIDD, OTS, RealRain-1k, LOL, and UIEB to target datasets such as Polyu, URHI, LHP-Rain, LIME, and UFO-120 (Fan et al., 28 May 2025). DATPRL-IR extends the generalization question further by covering natural, medical, and remote sensing domains within one model and reporting gains in both 6-task and 9-task multi-domain settings, while also showing zero-shot improvements on AIGC images, comics, specular highlight removal, desnowing, and RSI deblurring (Dong et al., 2 Mar 2026).

These results do not imply that AIR has solved all robustness problems. WeatherBench shows that even strong AIR architectures can behave unevenly across haze, rain, and snow, and that training on unified real data rather than mixed synthetic sources materially changes cross-domain outcomes (Guan et al., 15 Sep 2025). The field’s empirical trajectory therefore points not only toward stronger models, but toward stricter evaluation under real-world distribution shift.

6. Open problems and future directions

Several open directions recur across the literature. One is broader degradation coverage. OPIR explicitly suggests extending task-aware inverse operators to noise, motion blur, and low-light enhancement (Gao et al., 15 Jan 2026), while DATPRL-IR frames multi-domain AIR as only beginning with natural, medical, and remote sensing images and leaves additional domains such as industrial, microscopic, and artistic imagery open (Dong et al., 2 Mar 2026).

A second is better handling of unseen and mixed degradations. WeatherBench notes that it focuses on single weather degradation per image rather than mixed conditions such as rain+fog, and ClearAIR’s results on CDD-11 indicate that composite degradation remains a decisive benchmark for future AIR systems (Guan et al., 15 Sep 2025, Zhang et al., 6 Jan 2026). DAIR also notes artifacts in extremely severe low-light conditions and over-enhancement in some underwater scenes, which indicates that even strong latent-prior systems can fail under extreme distributional shift (Sharif et al., 22 Sep 2025).

A third direction is faster and more practical generative AIR. UniLDiff argues that latent diffusion priors are powerful but still limited by sampling speed and VAE compression, motivating faster diffusion sampling, better degradation estimation, and extension to video (Cheng et al., 31 Jul 2025). RestoreVAR shows one route around iterative denoising through visual autoregressive generation (Rajagopalan et al., 23 May 2025). FAPE-IR suggests a different route in which a frozen MLLM planner produces frequency-aware restoration plans and routes high- and low-frequency LoRA experts inside a diffusion executor, coupling semantic planning with frequency-based restoration (Liu et al., 18 Nov 2025). This suggests that AIR may increasingly split into a perception module and an execution module, rather than relying on a single monolithic backbone.

A fourth direction is stronger perceptual modeling. ClearAIR proposes integrating Just Noticeable Difference mechanisms into AIR, moving from a coarse imitation of Human Visual Perception toward perceptual-threshold-aware control (Zhang et al., 6 Jan 2026). Perceive-IR already shifts the discussion from degradation labels to fine-grained quality tiers in CLIP space (Zhang et al., 2024). A plausible implication is that future AIR systems will optimize not only reconstruction fidelity and task generality, but also a more explicit perceptual notion of when, where, and how much restoration should occur.

Across these developments, a common pattern is visible. AIR is moving away from static task tags and toward learned internal representations of degradation, spatially adaptive restoration policies, and evaluation protocols that stress real-world domain gap, compound degradation, and computational efficiency. The field now includes latent-prior inference, operator learning, cluster-guided routing, domain-aware prompting, barycenter-space decomposition, MLLM-guided planning, and HVP-inspired coarse-to-fine reasoning, all within the same broad objective: a single restoration system that remains accurate, efficient, and robust across heterogeneous degradations (Sharif et al., 22 Sep 2025, Gao et al., 15 Jan 2026).