All-in-One Image Restoration (AiOIR)

• AiOIR is a unified image restoration framework that recovers high-quality images from diverse unknown degradations by learning both degradation-specific and shared priors.
• It integrates multiple degradation types—including noise, blur, haze, rain, and low-light distortion—within one model, reducing complexity compared to separate-task approaches.
• Recent advances leverage tailored architectures such as prompt-based, routing-based, and one-step generative methods to achieve competitive restoration metrics with improved efficiency.

All-in-One Image Restoration (AiOIR) denotes a unified image restoration paradigm in which a single model recovers a clean image $x$ from an input $y$ degraded by unknown corruption, rather than using a separate network for each degradation type. In the unified degradation model, one assumes $y = D(x) + n$, where $D(\cdot)$ is a possibly unknown spatial degradation operator and $n$ is additive noise; AiOIR differs from traditional single-task restoration by explicitly integrating multiple degradations within one framework and by learning both degradation-specific and shared priors across corruptions such as noise, blur, haze, rain, snow, and low-light distortion (Jiang et al., 2024).

1. Problem formulation and operating regimes

AiOIR is commonly formalized as learning a single mapping that restores images degraded by one, or an unknown mixture, of multiple corruption types. EvoIR states the task as learning a single model $f_{\theta}$ such that, given $y$ degraded by one—or an unknown mixture—of additive Gaussian noise, haze, rain streaks, motion or defocus blur, or low-light distortion, the model recovers a high-quality image $x$ via $y = D(x)$ with $$D \in \{\text{Noise, Haze, Rain, Blur, Low-light}, \dots\}$$ (Ma et al., 4 Dec 2025). The survey further distinguishes AiOIR from Multiple-Degradation IR: the latter may train on a mixture of corruptions, but offers no architecture or learning design to specialize per-degradation, whereas AiOIR explicitly integrates multiple degradations within a unified framework (Jiang et al., 2024).

The field now spans several operating regimes. A common benchmark split is the “3D” setting, covering Noise+Haze+Rain, and the “5D” setting, which adds Blur+Low-light; EvoIR trains these settings by mixing data from BSD, RESIDE, Rain100L, GoPro, and LOL (Ma et al., 4 Dec 2025). More difficult protocols evaluate compound degradations, as in DAIR’s five compound settings and unseen compound settings, UniLDiff’s CDD11 evaluation for mixed degradations, RestoreVAR’s real-world generalization benchmarks, and ClearAIR’s CDD-11 experiments (Sharif et al., 22 Sep 2025). AiOIR has also expanded beyond natural-image single-domain restoration. DATPRL-IR defines Multi-Domain AiOIR as a single model spanning natural, medical, and remote-sensing domains, with tasks including natural-scene super-resolution, deraining, deblurring, CT denoising, MRI super-resolution, PET synthesis, remote-sensing super-resolution, cloud removal, and dehazing (Dong et al., 2 Mar 2026).

A recurring misconception is that AiOIR is simply “blind restoration with mixed-task training.” The literature treats it more narrowly: AiOIR methods are expected to incorporate degradation-aware routing, prompting, latent priors, clustering, domain adaptation, or other mechanisms that reconcile heterogeneous degradations within one parameterization rather than merely averaging over tasks (Jiang et al., 2024).

2. Architectural families and conditioning strategies

The survey organizes AiOIR methods along three axes: network architecture, learning strategy, and key improvements. Architecturally, the field includes task-specific encoders or decoders, shared decoder plus multiple heads, single-branch end-to-end models, shared encoder plus multiple decoders, pretrained mid-level backbones, Mixture-of-Experts, pretrained vision-language priors, visual prompting, textual or multimodal prompting, and question-answering formulations (Jiang et al., 2024). This taxonomy remains accurate for later methods, but newer work sharpens how degradation information is encoded and injected.

Family	Representative methods	Defining mechanism
Prompt-based	AutoDIR, DATPRL-IR, Cat-AIR	Visual, textual, or task/domain prompt injection
Expert/routing-based	DaAIR, ClusIR, FAPE-IR	Sparse or hierarchical expert activation
Frequency/wavelet-based	EvoIR, AnyIR, CWP-Net	Explicit spectral decomposition or wavelet disentangling
Latent-prior / physical-model based	DAIR, OPIR, D$y$0Net	Learned latent priors, inverse operators, or degradation decomposition

Prompt-centric methods treat degradation recognition as a conditioning problem. AutoDIR combines a Blind Image Quality Assessment stage, which parses unknown degradations into a text prompt via a fine-tuned CLIP, with a latent diffusion restoration stage and a Structural-Correction Module; it supports open-vocabulary user control and can auto-decompose multi-artifact images into sequential restoration steps (Jiang et al., 2023). DATPRL-IR further separates task and domain knowledge into two prompt pools, retrieves top-$y$1 task and domain prompts by cosine similarity, composes them into instance-level representations, and fuses them through cross-attention and Adaptive Gated Fusion to guide decoder layers in a multi-domain setting (Dong et al., 2 Mar 2026). Cat-AIR uses “prompt” modules in the decoder together with alternating spatial–channel attention, and its smooth learning strategy extends a 3D model to four and then five tasks while retaining original-task performance within $y$2 (Jiang et al., 23 Mar 2025).

Routing-based models seek conditional specialization without full model duplication. DaAIR uses a Degradation-aware Learner in the low-rank regime, combining a shared agnostic backbone, a pool of degradation-specific experts, and a self-learnable controller updated by EMA; at inference, each input patch is routed to exactly one specialist expert with top-1 sparse routing, while the whole model remains at $y$3 parameters and $y$4 MACs at $y$5 (Zamfir et al., 2024). ClusIR replaces flat MoE routing with a Probabilistic Cluster-Guided Routing Mechanism that factorizes degradation recognition and cluster-conditional expert activation, then propagates cluster priors into a Degradation-Aware Frequency Modulation Module (Hu et al., 11 Dec 2025). FAPE-IR uses a frozen MLLM planner to emit a frequency-aware plan and a LoRA-MoE executor that selects a high- or low-frequency expert through dual-end gating that merges text and spectral routing (Liu et al., 18 Nov 2025).

Other methods pursue degradation awareness through internal latent structure rather than explicit prompts. D$y$6Net employs a Cross-Domain Degradation Analyzer that fuses Fourier-domain degradation cues with spatial features to generate a Degradation Correction Prompt and a Strategy Prompt, which guide a Dynamic Decomposition Mechanism with Gumbel-Softmax-gated Adaptive Decomposition Blocks (Wang et al., 26 Feb 2025). DAIR reframes AIR as learned latent prior inference and structures restoration around three questions—“which features to route,” “where to restore,” and “what to restore”—implemented by adaptive feature selection, degradation maps, and FiLM-style semantic conditioning on a latent prior $y$7 (Sharif et al., 22 Sep 2025). OPIR, by contrast, adopts a physical degradation modeling perspective and predicts a task-aware inverse degradation operator, implemented as a spatially varying $y$8 kernel per pixel modulated by a Task-Aware Module (Gao et al., 15 Jan 2026).

3. Diffusion, autoregression, and single-step generative restoration

Generative AiOIR has been dominated by latent diffusion models, largely because they recover abundant texture details and exhibit strong generalization. AutoDIR is an early example: it modifies the latent diffusion U-Net by concatenating the degraded latent $y$9 to the noisy latent before each block, injecting text embeddings via cross-attention, and adding a Structural-Correction Module to repair faces and text that diffusion may distort (Jiang et al., 2023). UniLDiff advances this line by inserting a Degradation-Aware Feature Fusion module into early diffusion U-Net layers and a Detail-Aware Expert Module into the VAE decoder; it reports that, across dehaze, derain, denoise, deblur, and low-light tasks, its three-task setup is competitive in PSNR/SSIM with leading non-diffusion AiOIR methods while outperforming baselines in perceptual metrics such as LPIPS, DISTS, CLIP-IQA, MUSIQ, and MANIQA (Cheng et al., 31 Jul 2025).

A central practical limitation of diffusion-based AiOIR has been inference cost. Several 2025–2026 methods explicitly target one-step or near-one-step restoration. DOD builds on a pretrained Stable Diffusion model and is engineered to “restore once and stop.” It extracts degradation features from the pretrained SD U-Net at $y = D(x) + n$0 through Multi-degradation Feature Modulation, injects them into conditional LoRA updates, and adds a High-fidelity Detail Enhancement module in the decoder. DOD uses exactly $y = D(x) + n$1 denoising step versus $y = D(x) + n$2–$y = D(x) + n$3 steps in prior diffusion methods; on a single 3090 GPU with $y = D(x) + n$4 inputs, it runs in $y = D(x) + n$5 and is reported as $y = D(x) + n$6 faster than DA-CLIP, while achieving restoration quality of approximately $y = D(x) + n$7 and $y = D(x) + n$8 with LPIPS $y = D(x) + n$9 and DISTS $D(\cdot)$0 (Tang et al., 5 Aug 2025).

UDBM also targets single-step inference, but from a diffusion-bridge viewpoint. It replaces the strict terminal constraint of standard diffusion bridges with a relaxed Gaussian terminal whose variance is proportional to a pixel-wise uncertainty map, thereby keeping the score correction bounded as $D(\cdot)$1. It then introduces an uncertainty-guided noise schedule and path schedule and performs restoration in one pass. In a five-task AiO setting, UDBM-L reaches an average $D(\cdot)$2 of $D(\cdot)$3, runs at about $D(\cdot)$4 on $D(\cdot)$5, uses approximately $D(\cdot)$6, and is reported as a $D(\cdot)$7 speedup over DiffUIR (Tu et al., 29 Jan 2026). FAPE-IR likewise adopts a one-step latent diffusion setup, but its novelty lies in the planner–executor split: a frozen MLLM produces a plan $D(\cdot)$8, and a LoRA-MoE diffusion executor routes between two frequency experts. It is reported to run in approximately $D(\cdot)$9 per $n$0 image on an H200 GPU while achieving state-of-the-art performance across seven tasks (Liu et al., 18 Nov 2025).

Not all fast generative AiOIR remains diffusion-based. RestoreVAR replaces iterative denoising with visual autoregressive next-scale prediction over $n$1 scales, conditions the transformer with degraded-image latents through cross-attention, and adds a lightweight latent-space refinement transformer. On a single RTX 4090, RestoreVAR runs in $n$2 and $n$3, compared with Diff-Plugin at $n$4 and $n$5, AutoDIR at $n$6 and $n$7, and PixWizard at $n$8 and $n$9; it is described as the best generative method on PSNR/SSIM/LPIPS across dehazing, desnowing, deraining, low-light enhancement, and deblurring (Rajagopalan et al., 23 May 2025).

These results make clear that generative AiOIR no longer implies long diffusion chains. One line distills or restructures diffusion into one-step restorers; another abandons diffusion schedules in favor of autoregression while preserving generative priors. This suggests that the central design question has shifted from “whether to use generative modeling” to “how to inject degradation-aware guidance into a generative prior with acceptable latency.”

4. Frequency modeling, latent reasoning, and causal or physical structure

A prominent trend in AiOIR is the move from implicit degradation awareness to explicit structure in either the frequency domain, the latent space, or a causal or physical model. EvoIR exemplifies frequency-first design. Its Frequency-Modulated Module decomposes features into low- and high-frequency branches, applies learned spectral gating on the low-frequency branch and spatial gating on the high-frequency branch, and refines the fused result with residual Transformer blocks. It couples this with an Evolutionary Optimization Strategy that periodically re-optimizes the balance between fidelity and perceptual loss terms. In the 3-degradation setting, EvoIR reaches $f_{\theta}$0, outperforming DFPIR at $f_{\theta}$1 and AdaIR at $f_{\theta}$2; in the 5D setting it reaches $f_{\theta}$3 versus DFPIR at $f_{\theta}$4 and AdaIR at $f_{\theta}$5 (Ma et al., 4 Dec 2025).

AnyIR uses a different spatial–frequency decomposition. Its Gated Degradation Adaptation reweights three sub-latent components after an interleaved split of the feature channels, and its spatial-frequency parallel fusion combines spatial cross-enhancement with FFT-domain merging. Quantitatively, AnyIR reports $f_{\theta}$6 in the 3-degradation setting and $f_{\theta}$7 in the 5-degradation setting, while reducing model complexity by around $f_{\theta}$8 in parameters and $f_{\theta}$9 in FLOPs relative to PromptIR (Ren et al., 19 Apr 2025). D$y$0Net also relies on frequency cues, but in a cross-domain form: its CDDA analyzes the amplitude spectrum of the input, fuses it with spatial features through two cross-attention modules, and produces two prompts that guide a gated progressive decomposition process. On SOTS-Outdoor it reaches $y$1, which is reported as $y$2 over InstructIR at $y$3, and on GoPro deblurring it reaches $y$4, $y$5 over the prior best (Wang et al., 26 Feb 2025).

Wavelet formulations add another layer of structure. CWP-Net argues that AiOIR suffers from spurious correlation between non-degradation semantic features and degradation patterns, together with biased estimation of degradation patterns. Its encoder and decoder replace standard down- and up-sampling with wavelet attention modules, and its Wavelet Prompt Block constructs an alternative variable $y$6 for back-door adjustment, formalized as

$y$7

On the five-pattern setting, it reports a gain of $y$8 over the prior best average PSNR; on the seven-pattern setting, the gain is $y$9. It also reports $x$0 parameters, $x$1, and $x$2 inference time (Wang et al., 4 Mar 2026). ClusIR combines cluster semantics with wavelet-domain modulation: its PCGRM learns degradation clusters and cluster-conditional expert routing, while DAFMM uses those priors to modulate low- and high-frequency subbands. ClusIR reaches $x$3 in the three-task setting, $x$4 in the five-task setting, and $x$5 on CDD11 composite degradations (Hu et al., 11 Dec 2025).

Latent reasoning approaches criticize fixed spectral heuristics. DAIR explicitly states that hand-crafted methods, including frequency priors, assume known spectral signatures and often fail on mixed or unseen degradations. It instead learns a continuous latent prior from the input and organizes restoration around adaptive feature selection, spatial localization through degradation maps, and degradation semantics via FiLM-style modulation. It reports an average PSNR improvement of $x$6 and $x$7 efficiency, with $x$8 inference and $x$9 on $y = D(x)$0 images using an RTX-3060 (Sharif et al., 22 Sep 2025). OPIR offers a different form of structure: rather than regressing residuals, it predicts a task-aware inverse operator and an uncertainty perception map in a two-stage pipeline. In the all-in-one setting over rain, snow, and haze, it reaches an average $y = D(x)$1 of $y = D(x)$2, outperforming the best prior all-in-one method by $y = D(x)$3 PSNR, while running in $y = D(x)$4 versus $y = D(x)$5–$y = D(x)$6 for prompt- or large-model baselines (Gao et al., 15 Jan 2026).

These strands do not agree on a single privileged representation of degradation. Some methods claim explicit frequency modeling is essential; others argue that prompt-based or hand-crafted priors are brittle on mixed or unseen degradations. The literature therefore supports a narrower conclusion: explicit structure is useful, but the most effective structure may be frequency, latent, causal, or physical depending on the target degradation regime.

5. Training objectives, datasets, and benchmark culture

AiOIR training is typically multi-objective. The survey groups learning strategies into contrastive learning, multi-task learning, continual learning, and machine unlearning, and lists standard paired metrics such as PSNR and SSIM, perceptual metrics such as LPIPS and FID, and no-reference metrics such as NIQE, BRISQUE, CLIP-IQA, and MUSIQ (Jiang et al., 2024). Later methods instantiate these strategies in task-specific ways.

Loss design often reflects each method’s theory of task conflict. EvoIR uses

$y = D(x)$7

and re-optimizes $y = D(x)$8 with a population-based evolutionary process every $y = D(x)$9 iterations. Its ablation in the 3-degradation setting shows a Restormer-style baseline at $$D \in \{\text{Noise, Haze, Rain, Blur, Low-light}, \dots\}$$0; $$D \in \{\text{Noise, Haze, Rain, Blur, Low-light}, \dots\}$$1 only at $$D \in \{\text{Noise, Haze, Rain, Blur, Low-light}, \dots\}$$2; $$D \in \{\text{Noise, Haze, Rain, Blur, Low-light}, \dots\}$$3 only at $$D \in \{\text{Noise, Haze, Rain, Blur, Low-light}, \dots\}$$4; and full EvoIR at $$D \in \{\text{Noise, Haze, Rain, Blur, Low-light}, \dots\}$$5 (Ma et al., 4 Dec 2025). DOD uses a two-stage scheme: first it trains only LoRA matrices and a FiLM MLP with a reconstruction-plus-LPIPS data term and a KL-based distribution-matching distillation term, then it freezes PLA and trains only the RRDB-based HDE with reconstruction plus SSIM, using AdamW and LoRA rank $$D \in \{\text{Noise, Haze, Rain, Blur, Low-light}, \dots\}$$6 (Tang et al., 5 Aug 2025).

Some methods separate degradation recognition from restoration. AutoDIR trains BIQA with a multi-class fidelity loss plus a semantic-agnostic constraint, then trains AIR with denoising score matching and SCM with an $$D \in \{\text{Noise, Haze, Rain, Blur, Low-light}, \dots\}$$7 reconstruction objective in image space (Jiang et al., 2023). UDAIR trains source-domain restoration with an MAE loss plus Cross-Sample Contrastive Learning over codebook embeddings, then performs test-time adaptation on target images by minimizing a CORAL loss for five update steps (Fan et al., 28 May 2025). ClusIR uses $$D \in \{\text{Noise, Haze, Rain, Blur, Low-light}, \dots\}$$8 plus MS-SSIM with prototype normalization and orthogonality regularization (Hu et al., 11 Dec 2025). OPIR supervises both stages with Charbonnier, edge-aware Laplacian, and FFT-based frequency losses (Gao et al., 15 Jan 2026). ClearAIR combines an $$D \in \{\text{Noise, Haze, Rain, Blur, Low-light}, \dots\}$$9 reconstruction loss with an internal clue reuse consistency term, and trains its MLLM-based IQA module with an $y$00 quality regression loss (Zhang et al., 6 Jan 2026).

Benchmarking has also broadened. Standard paired datasets still dominate: BSD68 and Kodak24 for denoising, Rain100L or Rain13K for deraining, RESIDE or SOTS for dehazing, GoPro for deblurring, and LOL for low-light enhancement (Jiang et al., 2024). But recent AiOIR papers emphasize cross-setting evaluation. CDD11 is used for mixed degradations by UniLDiff, ClearAIR, AnyIR, and ClusIR (Cheng et al., 31 Jul 2025). DATPRL-IR introduces 6-task and 9-task, 3-domain evaluations across natural, medical, and remote-sensing datasets and reports DATPRL-IR at about $y$01 average in the 6-task setting, up $y$02 over MoCEIR (Dong et al., 2 Mar 2026). This suggests that benchmark design has become a methodological issue in its own right: method comparisons are increasingly sensitive to whether one evaluates single degradations, mixed degradations, unseen degradations, or cross-domain transfer.

6. Generalization, efficiency, and unresolved directions

A defining claim of AiOIR is generalization beyond closed, synthetic settings. Several methods report concrete evidence. AutoDIR shows zero-shot decomposition on Under-Display Camera and EUVP images, where BIQA proposes sequential steps such as “brightening” then “deblurring” or “deblurring” then “dehazing” (Jiang et al., 2023). D$y$03Net reports that Under-Display Camera generalization outperforms state of the art without fine-tuning (Wang et al., 26 Feb 2025). DAIR evaluates unseen compound degradations, where it reaches $y$04 versus PromptIR at $y$05 (Sharif et al., 22 Sep 2025). DATPRL-IR reports zero-shot gains on AIGC and Manga109 super-resolution, specular highlight removal, snow removal, real-world deraining, and remote-sensing deblurring, including $y$06 over MoCEIR on RSI deblurring with random blur (Dong et al., 2 Mar 2026). UDAIR specifically targets the gap between controlled source-domain training and real-world target-domain testing, using a codebook of degradation prototypes plus test-time CORAL adaptation, and reports state-of-the-art performance across ten datasets with $y$07 parameters and $y$08 FLOPs (Fan et al., 28 May 2025).

Efficiency has become equally central. DaAIR achieves $y$09 parameters and approximately $y$10 GMACs at $y$11, compared with PromptIR at $y$12 and IDR at $y$13, while reaching $y$14 in the three-degradation setting and $y$15 in the five-degradation setting (Zamfir et al., 2024). AnyIR reaches $y$16 parameters, $y$17 FLOPs, and $y$18 memory at $y$19, alongside $y$20 and $y$21 in the 3D and 5D settings (Ren et al., 19 Apr 2025). OPIR reports $y$22 inference with state-of-the-art all-in-one performance (Gao et al., 15 Jan 2026). DOD, RestoreVAR, UDBM, and FAPE-IR each show that strong perceptual restoration can be combined with one-step or near-one-step inference (Tang et al., 5 Aug 2025).

The open questions identified in the survey remain largely intact. These include task conflicts under joint optimization, out-of-distribution degradations, model complexity and efficiency, limited high-quality paired data, high-dimensional extensions to video or 3D, realistic composite degradations, and standardized super-benchmarks (Jiang et al., 2024). Later papers sharpen these directions rather than replacing them. DATPRL-IR points to hierarchical or dynamically pruned prompt pools and continual or domain-incremental restoration via prompt expansion without full retraining (Dong et al., 2 Mar 2026). RestoreVAR identifies stronger VQ-VAEs, improved refiner designs, and larger VAR backbones as key future levers (Rajagopalan et al., 23 May 2025). ClearAIR suggests that global quality assessment, region awareness, task recognition, and fine-detail reuse can be organized in an HVP-inspired “How → Where → What” order, which outperforms alternative orderings by more than $y$23 in its ablation (Zhang et al., 6 Jan 2026).

Taken together, the literature suggests that AiOIR is converging on a small set of durable principles: unified restoration requires explicit handling of degradation heterogeneity; blind restoration performance depends heavily on how degradation cues are represented and routed; and practical success increasingly demands simultaneous progress in fidelity, perceptual quality, efficiency, and transfer beyond the synthetic benchmarks on which the field was originally built.