Extreme Blind Image Restoration (EBIR)

• EBIR is a paradigm for blind image restoration handling severe, composite, and unseen degradations using strong priors and adaptive conditioning.
• It extends traditional blind restoration techniques to tasks like deraining, deblurring, and super-resolution by embracing zero-shot and training-free methods.
• Modern EBIR leverages diffusion-based models to decompose the restoration process, yielding improved perceptual quality and enhanced fidelity.

Extreme Blind Image Restoration (EBIR) denotes a stringent regime of blind image restoration in which the degraded observation is affected by unknown degradations that may be severe, compounded, out-of-distribution, or difficult to parameterize. In recent arXiv literature, EBIR is treated as a setting beyond ordinary blind image restoration: some papers emphasize fully blind and partially blind restoration without retraining or fine-tuning a diffusion prior, whereas others define EBIR through severely and compoundly degraded inputs that exceed the training scope of standard blind image restoration models. Across these formulations, the central difficulty is that both the clean image and the effective degradation process are underdetermined, so restoration must rely on strong image priors, internal degradation inference, adaptive conditioning, or explicit mechanisms for handling solution ambiguity (Chihaoui et al., 27 Mar 2025, Kim et al., 1 Oct 2025, Huang et al., 2024).

1. Scope, terminology, and problem setting

A useful working taxonomy in this literature distinguishes zero-shot IR, fully blind IR, and partially blind IR. In the zero-shot setting, no task-specific training on paired degraded/clean data is performed; a pre-trained model is used at test time. In the fully blind setting, both the degradation type and its parametric form or values are unknown. In the partially blind setting, the parametric form of the degradation is known, but its parameters are unknown. Representative fully blind tasks include deraining, JPEG de-artifacting, and raindrop removal; representative partially blind tasks include motion deblurring, Gaussian deblurring, and super-resolution (Chihaoui et al., 27 Mar 2025).

The broader EBIR literature extends this taxonomy by stressing unknown, unseen, and composite distortions. In this stronger sense, blind restoration is not merely kernel estimation with unknown parameters; it includes cases where the corruption may be mixed across types, spatially heterogeneous, or beyond the training distribution of conventional blind image restoration systems. Papers on adaptive blind all-in-one restoration and instant generative reference make this point explicitly by targeting unseen degradations, composite distortions, and severe real-world corruption rather than only single-degradation synthetic setups (Serrano-Lozano et al., 2024, Huang et al., 2024).

The underlying inverse problem is commonly written in terms of a degraded observation $y$, a clean image $x$, and an unknown forward operator such as $\mathcal H$ or $\mathcal D$. What distinguishes EBIR is not a new mathematical form so much as a stronger ill-posedness: the degradation operator may be unavailable, unreliable, or not expressible by a simple family, while the image prior must compensate for the resulting ambiguity. This is why recent EBIR methods either move restoration into latent spaces with known priors, estimate degradation representations internally, or decompose the restoration path into more tractable intermediate stages (Chihaoui et al., 27 Mar 2025, Tu et al., 2024, Kim et al., 1 Oct 2025).

2. Antecedents in blind restoration

EBIR emerged from several earlier lines of blind restoration research. One antecedent is blind image inpainting, where the system restores a clean image from a corrupted input without knowing the mask or the locations of missing pixels in advance. In that setting, the observation model is

$$x(i) = \begin{cases} y(i) + n(i), & M(i)=1, \ u(i), & M(i)=0, \end{cases}$$

and the blind aspect is precisely that the mask $M$ is not provided to the algorithm. The proposed solution combined an encoder-decoder structure, residual learning, pre-extracted gradient prior, and an $L_1$ reconstruction loss to handle outliers and arbitrary corruption geometry (Liu et al., 2017).

A second antecedent is blind universal restoration without prior knowledge, which argued that high-quality restoration requires degradation parameters to be supplied or estimated internally, but sought to remove explicit prior knowledge of the degradation process or its parameter space. Its Self-Normalization Side-Chain performed Global Weighted Average Pooling and multiplicative relevance modulation so that degradation-related parameters were deduced from the variety in the training data rather than from external labels or predefined parameterizations (Elron et al., 2020).

A third antecedent is classical blind deconvolution and regularization-based blind restoration, where the latent sharp image and blur kernel are estimated jointly. In the formulation

$g = f * h + n,$

the restoration objective combines a data term with priors on both the latent image and the blur kernel. An example is adaptively sparse regularization, which used first-order and second-order derivatives, an entropy-derived adaptive operator, Split Bregman optimization, and kernel positivity and normalization constraints to improve convergence and recovery accuracy under blur and noise (Xu, 2021).

A fourth antecedent is plug-and-play blind restoration with learned denoisers. CurvPnP treated the noise level as a variable rather than a fixed known quantity, alternating a data-fidelity update, blind noise estimation, and a denoising step conditioned on an estimated noise level and a Gaussian curvature map. This advanced blind restoration beyond non-blind Gaussian-denoising assumptions, but it still assumed Gaussian noise modeling and, for deblurring, a known blur kernel (Li et al., 2022).

3. Methodological families

The contemporary EBIR literature is heterogeneous, but several recurring design families can be identified.

Family	Representative papers	Characteristic mechanism
Classical blind optimization	(Xu, 2021, Li et al., 2022)	Joint estimation of image and degradation variables with explicit priors, iterative updates, and task-specific subproblems
Supervised blind restoration	(Liu et al., 2017, Elron et al., 2020)	End-to-end learning that infers corrupted regions or latent degradation states from paired degraded/clean data
Blind all-in-one restoration	(Qin et al., 2024, Serrano-Lozano et al., 2024)	One model for multiple degradations, using masked image modeling, segmentation heads, low-rank adapters, or adaptive merging
Diffusion-based blind restoration	(Lin et al., 2023, Tu et al., 2024, Xiao et al., 2024, Chihaoui et al., 27 Mar 2025, Huang et al., 2024, Kim et al., 1 Oct 2025, Eteke et al., 8 Sep 2025)	Pre-trained diffusion priors, training-free guidance, latent rectification, online degradation surrogates, generative references, or intermediate manifolds

Within this landscape, a recurring misconception is that any blind or all-in-one method is already operating in the EBIR regime. Recent papers argue otherwise. Adaptive blind all-in-one restoration states that standard blind all-in-one models usually require all possible degradation types to be defined during training and show limited generalization to unseen degradations. Restore Anything with Masks similarly emphasizes robustness across multiple degradations but is not framed as a dedicated extreme-degradation benchmark method. This literature therefore treats EBIR as stricter than ordinary blind multi-task restoration, especially when unseen or composite corruptions are central (Serrano-Lozano et al., 2024, Qin et al., 2024).

Another recurring distinction is between methods that explicitly model degradation and methods that avoid explicit degradation modeling. Classical blind deblurring, plug-and-play methods, and diffusion samplers with optimizable degradation surrogates remain close to the former. By contrast, latent-space rectification and instant-reference methods move toward the latter by letting the image prior dominate and by correcting the restoration trajectory in spaces where normality, semantic plausibility, or manifold alignment can be tested more directly (Xu, 2021, Tu et al., 2024, Chihaoui et al., 27 Mar 2025, Huang et al., 2024).

4. Diffusion priors and the modern EBIR paradigm

Diffusion models have become the dominant prior in modern EBIR because they provide a strong natural-image manifold while permitting test-time optimization or guidance. One influential direction decomposes blind restoration into degradation removal and information regeneration. DiffBIR first removes image-independent degradation with a restoration module and then regenerates lost detail with IRControlNet built on Stable Diffusion. Its region-adaptive restoration guidance constrains smooth regions more strongly than high-frequency regions, giving a controllable fidelity-realism trade-off through the guidance scale $s$ without retraining (Lin et al., 2023).

A related direction retains the diffusion prior but adds stronger guidance losses. Frequency-aware guidance for blind image restoration via diffusion models argued that spatial-domain consistency alone is insufficient and proposed a wavelet-based objective,

$$\mathcal{L}_{\text{freq}} = \left\Vert \mathbf{y} - \mathbf{\hat{y}} \right\Vert_2^2 + \sum_{i\in \{\text{LH, HL, HH}\}} \lambda_i \cdot \left\Vert \mathbf{y}_i - \hat{\mathbf{y}}_i \right\Vert_2^2,$$

so that restoration is constrained in both the spatial and frequency domains. It reported a 3.72 dB PSNR improvement over BlindDPS for motion blur and a 1.46 dB gain for Gaussian blur, while also improving FID and LPIPS (Xiao et al., 2024).

Another direction makes the degradation model itself adaptive during sampling. BIR-D used a pretrained DDPM as an image prior, an optimizable convolutional kernel and learnable mask as a degradation surrogate, and an empirical adaptive guidance scale computed in real time. The method updates the simulated degradation process and the restoration trajectory jointly, which is intended to support classical inverse problems, low-light enhancement, motion deblurring, HDR recovery, blind face restoration, and multi-degradation restoration within one framework (Tu et al., 2024).

Invert2Restore reformulated EBIR as noise inversion + noise rectification in the latent space of a pre-trained DDIM. The central observation is that when a degraded image is inverted through the diffusion model, the corresponding noise tends to lie in a low-density region of $x$0. Restoration is therefore performed by inverting the degraded image to $x$1, detecting local latent patches that fail a D’Agostino–Pearson omnibus test, replacing defective patches with nearest neighbors from Gaussian samples, and decoding the rectified latent. In the fully blind case the inversion is

$x$2

followed by patchwise normality testing and substitution. This design avoids explicit degradation modeling and directly corrects the input noise rather than an intermediate latent (Chihaoui et al., 27 Mar 2025).

InstantIR took a different approach by constructing an instant generative reference at each diffusion step. It used a DINO-based Degradation Content Perceptor, a Previewer trained by consistency distillation, and a Latent Aggregator that fuses the degraded input with the generated reference. The method also introduced adaptive restoration based on the variance behavior of these previews so that conditioning is amplified for higher-quality inputs and relaxed for highly degraded ones. The same framework supports text modulation for extreme degradation and creative restoration (Huang et al., 2024).

A more explicit EBIR formulation appears in prompt-conditioned information bottleneck approaches. These methods argue that direct ELQ-to-HQ restoration is unstable because the domain gap is too large, and instead learn a projector that maps an extreme low-quality image to an intermediate low-quality manifold that a frozen off-the-shelf blind image restoration model can handle. The objective combines low-quality reconstruction, high-quality prior matching, and a high-quality fidelity term, thereby recasting EBIR as a constrained manifold projection rather than a single intractable inverse map (Kim et al., 1 Oct 2025).

Low-complexity adapter designs constitute another branch of diffusion-based EBIR. BIR-Adapter injected degraded-image features extracted by the diffusion model itself into each self-attention layer, trained only lightweight additional weights, and added sampling guidance based on a low-frequency anchor from an initial restoration. Its experiments on synthetic cascades such as $x$3 and on RealSR targeted exactly the unknown and compounded degradations associated with EBIR, while emphasizing a substantially smaller parameter budget than ControlNet-style alternatives (Eteke et al., 8 Sep 2025).

5. Training regimes, benchmarks, and fidelity evaluation

Training regimes in EBIR range from fully supervised paired-data learning to zero-shot and training-free inference. Deep blind image inpainting, SNSC-based blind universal restoration, RAM, and ABAIR all rely on paired degraded/clean supervision, though they differ in how much degradation knowledge is externalized. Invert2Restore, Frequency-Aware Guidance, and several diffusion-guidance methods are explicitly training-free or zero-shot at inference, using a pre-trained diffusion prior without retraining or fine-tuning for the target degradation (Liu et al., 2017, Elron et al., 2020, Qin et al., 2024, Serrano-Lozano et al., 2024, Chihaoui et al., 27 Mar 2025, Xiao et al., 2024).

Experimental coverage is correspondingly broad. Recent papers report results on ImageNet 1K, CelebA 1K, the SPA validation set, FFHQ, ImageNet-1K validation, COCO, RESIDE, Cityscapes, GLD-based synthetic pretraining, RealSR, DRealSR, Rain100L, SOTS-Outdoor, BSD68, GoPro, LOL, and mixed-degradation suites such as blur + noise, blur + JPEG, and haze + snow. Common metrics include PSNR, SSIM, LPIPS, FID, MANIQA, MUSIQ, CLIP-IQA, LOE, PI, and Consistency, but the choice of metric depends strongly on whether the study prioritizes pixel fidelity, perceptual realism, or blind-restoration consistency (Lin et al., 2023, Serrano-Lozano et al., 2024, Huang et al., 2024).

Representative quantitative results illustrate the diversity of objectives rather than a single dominant benchmark. Invert2Restore reported the best results in its tables for both partially blind and fully blind tasks, including 22.56 PSNR / 0.666 SSIM / 0.347 LPIPS for motion deblur, 23.02 / 0.664 / 0.348 for Gaussian deblur, 22.62 / 0.671 / 0.330 for $x$4 super-resolution, and 27.12 / 0.809 / 0.355 for deraining, with a runtime of 190 s and memory of 1.2 GB on CelebA $x$5. ABAIR reported 31.17 dB / 0.921 in the five-task setup and 33.07 dB / 0.916 in the three-task setup, along with strong gains on unseen datasets and mixed degradations. Frequency-aware guidance reported the PSNR gains over BlindDPS already noted above, especially in blind deblurring (Chihaoui et al., 27 Mar 2025, Serrano-Lozano et al., 2024, Xiao et al., 2024).

Evaluation itself has become a controversy. A central claim of the Consistency with Degraded Image framework is that full-reference IQA metrics such as PSNR, SSIM, LPIPS, and DISTS can be misleading in blind restoration because of solution non-uniqueness and degradation indeterminacy. CDI therefore defines fidelity in terms of degraded-image consistency rather than direct agreement with a single reference. Its Reference Guided CDI operates in the wavelet domain through Wavelet Noise Splitting, Wavelet Noise Equivalent Attenuation, and Wavelet Attenuation Matching, while Reference Agnostic CDI learns a Wavelet Attenuation Extraction Net so that blind-restoration fidelity can be estimated without reference images. DISDCD, a Degraded Images Switch Display Comparison Dataset using a 2AFC protocol, was introduced to validate this claim; RGCDI achieved the highest 2AFC agreement, and RACDI was slightly below RGCDI (Tang et al., 24 Jan 2025).

6. Misconceptions, limitations, and open directions

Several misconceptions recur in the EBIR literature. First, blind does not imply assumption-free. Many methods described as blind still require a known degradation family, a parametric operator model, explicit parameter supervision, or iterative parameter correction. This is why Invert2Restore contrasts itself with BlindDPS, GibbsDDRM, BIRD, GDP, and DreamClean, and why SNSC emphasized that earlier blind techniques necessitated prior knowledge of the degradation process and assumptions regarding its parameter-space (Chihaoui et al., 27 Mar 2025, Elron et al., 2020).

Second, all-in-one does not imply extreme-blind generalization. ABAIR makes this explicit by noting that conventional blind all-in-one models usually require all possible degradation types to be defined during training and may generalize poorly to unseen distortions. Its own limitations remain significant: the model still depends on a predefined adapter set for seen degradations, new corruption types require training a new adapter and updating the estimator, the estimator is lightweight but not perfect, and composite degradations are addressed through adapter blending only among known adapters (Serrano-Lozano et al., 2024).

Third, perceptual realism and reference fidelity should not be conflated. Diffusion-based systems such as DiffBIR, InstantIR, and other generative restorers may sacrifice PSNR or SSIM while improving CLIP-IQA, MUSIQ, MANIQA, or FID. CDI formalizes the reason: in EBIR, the restored output is often not uniquely determined by the degraded observation, so a visually plausible image can be a valid solution even when full-reference metrics are unfavorable (Lin et al., 2023, Huang et al., 2024, Tang et al., 24 Jan 2025).

The current literature also exposes practical limits. DiffBIR requires about 50 diffusion sampling steps per image. Invert2Restore is broader in applicability than DreamClean but slower in runtime. Prompt-conditioned information bottleneck methods note that the frozen restoration backbone must be relatively simple for practical gradient flow, and prompt extraction from severely degraded inputs is not always ideal. These limitations indicate that computational cost, routing uncertainty, and stable conditioning remain core bottlenecks (Lin et al., 2023, Chihaoui et al., 27 Mar 2025, Kim et al., 1 Oct 2025).

A plausible implication is that future EBIR research will continue to converge on three themes already visible in the literature: stronger open-set degradation recognition and uncertainty-aware routing; restoration objectives and benchmarks that separate degraded-image consistency from perceptual quality; and modular architectures that project extreme low-quality inputs into manifolds where pretrained priors are already competent. That direction is consistent with the current movement from explicit degradation fitting toward latent rectification, adaptive guidance, instant references, and intermediate-manifold projection (Serrano-Lozano et al., 2024, Tang et al., 24 Jan 2025, Chihaoui et al., 27 Mar 2025, Kim et al., 1 Oct 2025).