Domain-Invariant Visual Restoration (DIVER)

• Domain-Invariant Visual Enhancement and Restoration (DIVER) is a unified framework that disentangles content from diverse distortions to robustly restore images across varying degradation types.
• It leverages causal modeling, vision-language alignment, diffusion strategies, and transformer-based architectures to generalize restoration performance on unseen degradations.
• Empirical studies demonstrate that DIVER frameworks achieve superior metrics (e.g., PSNR, SSIM) on tasks such as denoising, deblurring, deraining, and underwater enhancement.

Domain-invariant visual enhancement and restoration (DIVER) encompasses a body of methodologies designed to robustly improve the quality of images subject to diverse, often unknown, degradation types, without prior knowledge of the domain or the need for retraining or reconfiguration per domain. The DIVER paradigm seeks representations or model architectures that generalize across distortion types and acquisition conditions by explicitly disentangling content from degradation, aligning or learning invariant features, or otherwise unifying visual restoration in ways that mitigate domain shift and ensure high restoration fidelity across heterogeneous environments.

1. Causal and Representation-Invariant Foundations

The causal perspective on DIVER formulates image degradations as confounders that corrupt the direct relationship between the observed (distorted) image and the underlying clean image. In "Learning Distortion Invariant Representation for Image Restoration from A Causality Perspective" (Li et al., 2023), this is formalized via the structural equation:

$Y = g(X, D)$

where $X$ is the clean image, $D$ is the distortion (type or degree), and $Y$ is the degraded observation. The restoration network predicts $Ŷ$ from $Y$, but due to the causal path $D \rightarrow Y \leftarrow X$ and $D \rightarrow Ŷ$, there exists a backdoor path introducing spurious correlation.

To address this, the DIVER training objective leverages Pearl's back-door adjustment:

$$P(\hat{Y} \mid \text{do}(Y)) = \sum_{d \in \mathcal D} P(\hat{Y} \mid Y, D=d) P(D=d)$$

Practically, for a finite set of distortions, this is instantiated through meta-learning:

• Inner loop: Adapt the model on synthetic data with each sampled distortion $d_i$ as intervention.
• Outer loop: Aggregate updates to learn a representation invariant w.r.t. the distortion set.

Counterfactual distortion augmentation plays a crucial role: all distortions are synthetically generated to cover the support of $D$, and the restoration model is forced, through "virtual" model updating, to minimize reconstruction loss across these conditions. Substantial generalization gains to unseen (out-of-support) distortion types and degrees are observed in denoising, deblurring, hybrid, and real-world image restoration (Li et al., 2023).

2. Vision-Language, Contrastive, and Scene-Conditioned Models

Vision-language pretraining (CLIP) and scene-conditioned architectures form a new class of DIVER models in "VL-UR: Vision-Language-guided Universal Restoration" (Liu et al., 11 Apr 2025). The VL-UR system comprises three tightly integrated components:

• Zero-shot CLIP Backbone: Aligns visual and linguistic representations in a 512-dimensional shared space. The pretrained CLIP is partially adapted as a frozen backbone.
• Scene Classifier: Assigns each degraded image to one of eleven degradation types using cosine similarity between image and templated text embeddings ("The image has ⌈type⌉ degradation").
• Scene Restorer: A U-Net encoder-decoder, equipped with Cross-Transformer Aggregation (CTransAgg) blocks, leverages prompt-guided cross-attention (PGCA), injecting language-derived embeddings to guide restoration at multiple scales.

The model is trained with a composite objective:

$$L_\text{total} = γ_1 L_{\text{Smooth}ℓ_1} + γ_2 L_{\text{MS-SSIM}} + γ_3 L_{\text{CDRL}}$$

with terms penalizing structural, perceptual, and high-level feature discrepancies against ground truth.

Robust domain-invariant restoration is achieved by jointly learning visual-semantic alignment over eleven composite weather-induced degradations—e.g., haze, rain, snow, and their intersections—using the Composite Degradation Dataset (CDD-11). VL-UR achieves or ties for state-of-the-art PSNR/SSIM across all degradation types, outperforming both "one-to-one" and "all-in-one" restoration baselines without task-dependent switching or retraining (Liu et al., 11 Apr 2025).

3. Diffusion-Based and Multi-Domain Adaptation Paradigms

Diffusion models have been harnessed to explicitly bridge synthetic-real domain gaps and unify restoration across modalities. "Unified Image Restoration and Enhancement: Degradation Calibrated Cycle Reconstruction Diffusion Model" (CycleRDM) (Xue et al., 2024) exemplifies this approach:

• Two-stage conditional diffusion: Stage 1 maps the degraded input to a rough normal image; Stage 2 refines this to a high-fidelity output.
• Wavelet-domain calibration: Post-diffusion, the low-frequency wavelet subband is further denoised via conditional diffusion, targeting global consistency, while a feature gain module (CNN + residual dense blocks) corrects high-frequency detail.
• Multimodal and Fourier Guidance: CLIP-based semantic losses and frequency-aware spectral losses regularize both restoration structure and perceptual quality.

CycleRDM demonstrates strong generalization on blind enhancement tasks (low-light, underwater, backlight) and multiple linear restoration tasks (deblurring, denoising, deraining, inpainting), outperforming recent unified baselines in PSNR, SSIM, LPIPS, and FID on nine diverse benchmarks, despite being trained with minimal supervision (≤500 samples per task) (Xue et al., 2024).

Similarly, "Denoising as Adaptation: Noise-Space Domain Adaptation" employs a U-Net diffusion model not for inference, but as a denoising regularizer in the noise space, supervising the restoration network by ensuring both synthetic and real restoration outputs are consistent with a shared clean manifold. Critical techniques, such as channel shuffling and residual-swapping contrastive loss, force the network to generalize across both domains and prevent shortcut learning. Substantial improvements in real-world denoising, deblurring, and deraining confirm the efficacy of this approach (Liao et al., 2024).

4. Multi-Domain Transformers and Unification Across Degradations

Transformer-based architectures have been extended to multi-domain, domain-invariant visual restoration in "Image Restoration via Multi-domain Learning" (SWFormer) (Jiang et al., 7 May 2025). The central principle is the exploitation of shared spatial, wavelet, and Fourier priors:

• Spatial-Wavelet-Fourier Mixer (SWFM): Each block decomposes input features into spatial, wavelet, and Fourier streams, capturing local-to-global degradations with dedicated mechanisms.
• Multi-Scale ConvFFN (MSFN): Merges representations across spatial resolutions, further promoting consistency across structure scales.
• Lossless Multi-Input Multi-Output (LMIMO): Inputs and reconstructs at multiple scales via 2D wavelet transforms, ensuring robustness to spatially varying corruptions.

Supervision is imposed by a composite loss over spatial, wavelet, and frequency domains:

$$L^i = \|O^i - G^i\|_1 + \|W(O^i) - W(G^i)\|_1 + \lambda\|\mathcal{F}(O^i) - \mathcal{F}(G^i)\|_1$$

where $O^i$ and $G^i$ represent output and ground-truth at each wavelet scale; $W$ and $\mathcal{F}$ are learnable wavelet and FFT transforms.

Experimental results across ten restoration tasks—spanning haze, snow, blur, rain(rain-streak and raindrop), shadow, cloud, underwater, and low-light—on 26 datasets demonstrate SWFormer achieves top-tier PSNR/SSIM, with parameter/FLOP efficiencies and real-time inference not achievable by monolithic transformer designs (Jiang et al., 7 May 2025).

5. Domain-Invariant Restoration in Unsupervised and Physics-Guided Regimes

In unsupervised and physically-constrained settings, DIVER frameworks have advanced underwater image restoration. "Development of Domain-Invariant Visual Enhancement and Restoration (DIVER) Approach for Underwater Images" (Makam et al., 30 Jan 2026) integrates:

• IlluminateNet and Spectral Equalization Filter: Preprocess luminance and spectral balance, branching based on per-channel illumination conditions.
• Adaptive Optical Correction Module (AOCM): Channel-adaptive contrast and hue refinement, including hue speckle suppression in the CIELab domain.
• Hydro-OpticNet: Physics-guided modules for backscatter (VeilNet) and wavelength-dependent attenuation (AttenNet), incorporating depth priors and softplus-regularized functional forms.

Optimization employs composite loss terms (gray-world, luminous, Huber, color-consistency, Sobel gradient) in a fully unsupervised manner, using no paired ground-truth. In evaluation across eight diverse underwater datasets (ranging from shallow/turbid to deep and artificially illuminated), DIVER exhibits at least a 9% improvement in UCIQE over other state-of-the-art methods, the lowest GPMAE for color accuracy, and substantial gains in downstream task performance (ORB keypoint repeatability/matching) (Makam et al., 30 Jan 2026). Ablation confirms that empirical–physics hybrid modeling, in conjunction with adaptive loss design, is critical for strong domain-invariance.

6. Unsupervised Learning, Adversarial, and Consistency Constraints

Early DIVER schemes, such as in "Learning Invariant Representation for Unsupervised Image Restoration" (Du et al., 2020), implement a dual-domain adversarial framework, explicitly disentangling invariant (content/texture) and variant (noise) components via separate encoders. Restoration leverages adversarial alignment both in feature space (with a representation discriminator) and image space (via PatchGAN discriminators).

Self-supervised modules—including background and semantic consistency—ensure that not only foreground details but also scene context and semantics persist through translation and restoration. The total objective comprises adversarial, cycle consistency, self-reconstruction, background/semantic consistency, and KL regularization terms, all optimized in a min–max fashion.

Empirical studies on synthetic/real denoising and low-dose CT tasks reveal that DIVER achieves comparable or superior PSNR/SSIM to supervised and unsupervised baselines, while also realizing stable, rapid convergence (Du et al., 2020).

7. Limitations, Evaluation, and Directions for Extension

DIVER models fundamentally depend on the comprehensiveness of the synthetic degradation set ($D$) for maximal generalization (Li et al., 2023), and may require retraining or augmentation to handle "novel" domains or composite degradations not represented during optimization (Liu et al., 11 Apr 2025). Physically inspired models, while robust in deployment, require accurate environmental priors or depth estimation, and diffusion/proxy-based adaptation schemes involve computational overhead during joint training (Liao et al., 2024, Xue et al., 2024, Makam et al., 30 Jan 2026).

Potential extensions include learnable prompt adapters for prompt-based systems, front-door style interventions via latent proxies, continual learning for explicit adaptation, and integration with generative or diffusion-based priors for improved high-frequency fidelity (Liu et al., 11 Apr 2025, Xue et al., 2024, Li et al., 2023). The general aim is maximally domain-invariant restoration: a single, scalable model, insensitive to both distortion distribution and task-specificity, capable of robustly restoring images under real-world, dynamic, and previously unseen conditions.

---

Key domain-invariant visual enhancement and restoration frameworks are referenced in:

• "Learning Distortion Invariant Representation for Image Restoration from A Causality Perspective" (Li et al., 2023)
• "VL-UR: Vision-Language-guided Universal Restoration" (Liu et al., 11 Apr 2025)
• "Unified Image Restoration and Enhancement: Degradation Calibrated Cycle Reconstruction Diffusion Model" (Xue et al., 2024)
• "Denoising as Adaptation: Noise-Space Domain Adaptation for Image Restoration" (Liao et al., 2024)
• "Image Restoration via Multi-domain Learning" (Jiang et al., 7 May 2025)
• "Development of Domain-Invariant Visual Enhancement and Restoration (DIVER) Approach for Underwater Images" (Makam et al., 30 Jan 2026)
• "Learning Invariant Representation for Unsupervised Image Restoration" (Du et al., 2020)