Skip to content

ClearAIR: A Human-Visual-Perception-Inspired All-in-One Image Restoration

Conference: AAAI 2026 arXiv: 2601.02763 Code: Not released Area: Image Restoration Keywords: All-in-One Image Restoration, Human Visual Perception, MLLM-based IQA, Semantic Guidance, Self-supervised Learning

TL;DR

Inspired by human visual perception (HVP), this paper proposes ClearAIR, a coarse-to-fine unified image restoration framework that progressively recovers image quality through four stages — MLLM-based quality assessment → semantic region perception → degradation type identification → internal clue reuse — achieving state-of-the-art performance across multiple degradation tasks.

Background & Motivation

Poor generalization of task-specific models: Early image restoration methods design dedicated networks for individual degradations (denoising, dehazing, deraining, etc.), making cross-task generalization infeasible and incurring high deployment costs.

General models still require multiple instances: Although general restoration models such as NAFNet and Restormer can handle multiple degradation types, each degradation still requires a separate model instance, leading to complex inference pipelines.

Existing AiOIR methods overlook spatially non-uniform degradation: All-in-One methods such as AirNet and PromptIR apply uniform processing strategies over the entire image, ignoring spatial variations in degradation distribution and severity.

Texture complexity affects restoration difficulty: Even when degradation is uniformly distributed, flat regions and complex texture regions present significantly different restoration challenges, and a uniform strategy leads to over-smoothing or artifacts.

Lack of hierarchical perceptual mechanisms: Human vision processes scenes globally before locally, whereas existing methods lack a progressive perceptual pipeline from global structure to local detail.

Insufficient fine detail recovery: Existing methods remain limited in recovering fine-grained textures and fail to exploit the intrinsic structural information within images.

Method

Overall Architecture

ClearAIR emulates the hierarchical processing pipeline of human visual perception and consists of four core components:

  1. MLLM-based IQA (global quality assessment) → 2. SGU (region-level semantic perception) → 3. Task Identifier (degradation type recognition) → 4. ICRM (internal clue reuse)

The restoration backbone is Restormer, with Prompt Transformer Block counts of [3, 5, 6, 8] across four levels and channel dimensions of [48, 96, 192, 384].

Key Design 1: MLLM-based Overall Assessment

  • Employs DeQA (an MLLM-based image quality assessment model) as the global quality perceiver.
  • A visual encoder encodes the degraded image into visual tokens, which are compressed by a vision abstractor and fused with text tokens before being fed into the MLLM.
  • The hidden state \(\mathcal{Q}\) from the layer preceding the "quality level" token is extracted as the quality representation.
  • A Quality Guidance Module (QGM) injects this representation into the restoration backbone via affine transformation: \(\mathbf{X}_{out} = \mathbf{X}_{in} \odot \text{Linear}(\mathbf{F}_q) + \text{Linear}(\mathbf{F}_q)\)

Key Design 2: Semantic Guidance Unit (SGU) + Task Identifier

Region-level perception (SGU): - Pretrained SAM2 is applied to the degraded image to generate \(N_m\) binary semantic masks. - Mask Average Pooling (MAP) computes the mean features within each mask region and broadcasts them back to the corresponding positions, yielding a semantic prior \(\mathbf{F}_{sem}\). - A mask dropout strategy is introduced during training, randomly removing a subset of masks and merging them into the background to improve robustness. - Semantic features interact with the backbone via Semantic Cross-Attention (SCA).

Degradation recognition (Task Identifier): - DA-CLIP generates a content embedding \(\mathbf{F}_c \in \mathbb{R}^{1 \times 512}\) and a degradation embedding \(\mathbf{F}_d \in \mathbb{R}^{1 \times 512}\). - The degradation embedding is passed through an MLP + Softmax to produce weighted combinations of a learnable prompt set \(\mathcal{P}\), yielding the degradation prompt \(\mathbf{F}_p\). - A Degradation-Aware Module (DAM) performs content-aware enhancement via cross-attention using \(\mathbf{F}_c\), while \(\mathbf{F}_p\) generates a degradation mask for spatial feature modulation.

Key Design 3: Internal Clue Reuse Mechanism (ICRM)

  • Weak and strong augmentations are sequentially applied to the restored output \(\mathbf{I}_r\).
  • The L2 distance between the weakly and strongly augmented outputs serves as an internal consistency loss: \(\mathcal{L}_{inter} = \gamma \cdot \|\mathbf{I}_r^w - \mathbf{I}_r^s\|_2^2\)
  • This self-supervised approach mines intrinsic structural information within the image to enhance detail recovery.
  • No additional annotations are required; the mechanism exploits the internal statistics of the image itself.

Loss & Training

  • Total loss: \(\mathcal{L}_{total} = \mathcal{L}_1 + \alpha \cdot \mathcal{L}_{inter}\), where \(\alpha = 0.25\) and \(\gamma = 0.05\).
  • Optimizer: AdamW (\(\beta_1=0.9\), \(\beta_2=0.999\)), learning rate \(2 \times 10^{-4}\), batch size 4.
  • Training for 300K iterations; inputs randomly cropped to 256×256 with random horizontal/vertical flipping.
  • Hardware: NVIDIA RTX 4090.

Key Experimental Results

Three Degradations (Denoising + Dehazing + Deraining)

Method Params SOTS (PSNR/SSIM) Rain100L BSD68 σ=15 BSD68 σ=25 BSD68 σ=50 Avg
PromptIR 36M 30.58/.974 36.37/.972 33.98/.933 31.31/.888 28.06/.799 32.06/.913
AdaIR 29M 31.06/.980 38.64/.983 34.12/.934 31.45/.892 28.19/.802 32.69/.918
VLU-Net 35M 30.71/.980 38.93/.984 34.13/.935 31.48/.892 28.23/.804 32.70/.919
ClearAIR 31M 31.08/.981 38.61/.984 34.18/.935 31.50/.891 28.31/.804 32.74/.919

Five Degradations (+ Deblurring + Low-light Enhancement)

Method SOTS Rain100L BSD68 σ=25 GoPro LOLv1 Avg
Perceive-IR 28.19/.964 37.25/.977 31.44/.887 29.46/.886 22.81/.833 29.84/.909
AdaIR 30.53/.978 38.02/.981 31.35/.888 28.12/.858 23.00/.845 30.20/.910
ClearAIR 30.12/.978 38.20/.982 31.53/.888 29.67/.887 22.83/.846 30.45/.916

All-Weather (Snow + Rain-Haze + Raindrops)

Method Snow100K-S Snow100K-L Outdoor-Rain RainDrop Avg
Histoformer 37.41/.966 32.16/.926 32.08/.939 33.06/.944 33.68/.945
ClearAIR 37.79/.967 32.53/.932 32.45/.941 32.82/.942 33.90/.946

Composited Degradations (CDD-11)

ClearAIR achieves 29.34 dB / 0.886 SSIM, surpassing OneRestore (28.72 dB) by 0.62 dB.

Ablation Study

  • Perception order: "How-Where-What" (ours) achieves the best performance at 38.21 dB, outperforming "What-How-Where" (38.04) and "Where-What-How" (37.89).
  • Component contributions: Removing any one of IQA, SGU, TI, or ICRM leads to performance degradation; all four components together achieve the best result of 38.21/0.986.

Highlights & Insights

  1. Novel HVP-inspired progressive design: The four-stage pipeline adapts the human "global-before-local" visual perception principle to AiOIR, yielding a logically coherent framework.
  2. MLLM-driven quality assessment: This work is among the first to incorporate MLLM-based IQA as a global prior for image restoration, leveraging cross-modal understanding to enrich degradation representations.
  3. Region-level adaptive processing: The combination of SGU and Task Identifier enables differentiated handling of spatially non-uniform degradation, addressing a fundamental limitation of uniform processing strategies.
  4. Annotation-free self-supervised detail recovery via ICRM: The mechanism cleverly exploits intrinsic image statistics, using augmentation consistency constraints to enhance texture recovery without additional labels.
  5. Comprehensive SOTA across four AiOIR settings: With a moderate parameter count of 31M, ClearAIR consistently outperforms recent methods including AdaIR and VLU-Net.

Limitations & Future Work

  1. Inference efficiency concerns: The framework incorporates three large pretrained models — DeQA, SAM2, and DA-CLIP — which likely incur substantial inference overhead and latency; the paper does not report inference speed.
  2. Heavy reliance on pretrained models: The framework's performance is strongly dependent on the quality of DeQA, SAM2, and DA-CLIP, and their behavior under severe degradation is not sufficiently discussed.
  3. Limited gain from ICRM: In the ablation study, removing ICRM results in only a 0.18 dB drop (38.03→38.21), suggesting that the design complexity is not fully justified by the performance gain.
  4. Moderate performance on low-light enhancement: In the five-degradation setting, ClearAIR achieves 22.83 dB on LOLv1, falling short of AdaIR's 23.00 dB, indicating limited benefit from global quality guidance for low-light scenarios.
  5. Lack of large-scale real-world evaluation: Validation is primarily conducted on synthetic datasets; generalization to real-world degradation scenarios remains to be further verified.
  • AirNet (CVPR'22): Contrastive learning encodes degradation representations; ClearAIR extends this with region-level perception.
  • PromptIR (NeurIPS'23): Visual prompts guide multi-degradation processing; ClearAIR generates degradation prompts explicitly via DA-CLIP rather than relying purely on learnable parameters.
  • Perceive-IR (TIP'25): Jointly perceives degradation type and severity; ClearAIR further introduces MLLM-based global assessment and semantic region segmentation.
  • AdaIR (ICLR'25): Adaptive image restoration; ClearAIR surpasses it by 0.25 dB in average PSNR on the five-degradation setting.
  • VLU-Net (CVPR'25): Vision-language unified restoration; ClearAIR outperforms it by 0.04 dB on three degradations with fewer parameters.
  • OneRestore (ECCV'24): Composite degradation restoration; ClearAIR surpasses it by 0.62 dB on CDD-11.

Rating

  • Novelty: ⭐⭐⭐⭐ — The HVP-inspired four-stage framework is conceptually novel, and incorporating MLLM-based IQA into restoration is an innovative contribution.
  • Experimental Thoroughness: ⭐⭐⭐⭐ — Four AiOIR settings, ablation studies, and qualitative comparisons provide broad coverage.
  • Writing Quality: ⭐⭐⭐⭐ — The paper is well-structured, the HVP analogy is consistently maintained throughout, and mathematical expressions are presented rigorously.
  • Value: ⭐⭐⭐⭐ — The work advances the research direction of region-level adaptive perception in the AiOIR field.