RIRF: Reasoning Image Restoration Framework

Abstract: Universal image restoration (UIR) aims to recover clean images from diverse and unknown degradations using a unified model. Existing UIR methods primarily focus on pixel reconstruction and often lack explicit diagnostic reasoning over degradation composition, severity, and scene semantics prior to restoration. We propose Reason and Restore (R&R), a novel framework that integrates structured Chain-of-Thought (CoT) reasoning into the image restoration pipeline. R&R introduces an explicit reasoner, implemented by fine-tuning Qwen3-VL, to diagnose degradation types, quantify degradation severity, infer key degradation-related factors, and describe relevant scene and object semantics. The resulting structured reasoning provides interpretable and fine-grained diagnostic priors for the restorer. To further improve restoration quality, the quantified degradation severity produced by the reasoner is leveraged as reinforcement learning (RL) signals to guide and strengthen the restorer. Unlike existing multimodal LLM-based agentic systems that decouple reasoning from low-level vision tasks, R&R tightly couples semantic diagnostic reasoning with pixel-level restoration in a unified framework. Extensive experiments across diverse UIR benchmarks demonstrate that R&R achieves state-of-the-art performance while offering unique interpretability into the restoration process.

• The paper introduces a novel chain-of-thought diagnostic phase that detects and quantifies diverse degradations to inform targeted restoration.
• It employs fine-tuned vision-language models with reinforcement learning, achieving significant improvements in PSNR and SSIM across multiple benchmarks.
• The framework enhances restoration interpretability and robustness in adverse conditions, setting a new standard for explainable image restoration.

Reasoning Image Restoration via Chain-of-Thought: A Technical Analysis of RIRF

Introduction and Context

Universal Image Restoration (UIR) aims to recover clean images from input data containing diverse, unknown, or mixed degradations. Traditional pipelines address this task through direct pixel-level reconstruction—frequently either with models specialized for single degradations or via prompt-based generation that leaves detailed analysis of degradation composition to the user. Existing multimodal LLM (MLLM) approaches tend to segregate semantic reasoning (to diagnose scene or degradation) from the low-level pixel restoration, leading to suboptimal information flow.

The "Reason and Restore" (R&R) framework (2604.09511) introduces an explicit, structured Chain-of-Thought (CoT) reasoning phase prior to restoration, powered by fine-tuned vision-LLMs (Qwen3-VL and Qwen-Image-Edit). This diagnostic step detects and quantifies underlying degradations and feeds structured priors—grounded in both physical degradation parameters and semantic scene descriptors—into the restorer. A reinforcement learning stage, leveraging these degradation severity metrics as rewards, further aligns the restoration process with the diagnosed priors.

Methodology

Structured Diagnostic Reasoning (Reason Phase)

The reasoning module, fine-tuned Qwen3-VL, performs a diagnostic protocol:

• Degradation identification: Detects presence of multiple degradation types such as fog, blur, rain streaks, and sensor noise.
• Severity quantification: Assigns continuous scores (0–100) to each detected degradation, enabling prioritization during restoration.
• Parameter estimation: Infers physically and statistically meaningful parameters (e.g., atmospheric light for fog, kernel length/direction for blur, noise statistics).
• Scene-level semantics: Produces concise, clean-scene geometric and contextual descriptions, enforcing content-aware restoration.

Training data is generated using a compositional, semi-realworld degradation model. The CoT output format is strictly regularized (binary indicators, continuous scores, parameter estimates), maximizing interpretability and downstream utility.

Pixel-Level Restoration (Restore Phase)

Restoration is executed by Qwen-Image-Edit, a VLM-based universal restorer agnostic to specific degradation types. The model is explicitly conditioned on the entire diagnostic output of the Reason phase:

• Diagnostic cues are injected as natural language prompts in training and inference.
• Restoration is sensitive to diagnosed degradation composition, severity, and scene context, allowing the model to dynamically adjust restoration strategies for mixed or complex scenarios.
• Conditioning on explicit, structured priors (rather than generic free-text prompts) avoids over- or under-restoration and model hallucination in unseen contexts.

Reinforcement Learning with Diagnostic Rewards

Supervised fine-tuning is augmented with a reinforcement learning (RL) phase. Here, the reduction in degradation severity—predicted by the diagnostic reasoner after restoration—serves as a reward, closing the loop between high-level analysis and pixel-level restoration.

• A group relative policy optimization (GRPO) strategy is employed, leveraging a fidelity-based policy proxy for RL updates.
• The reward aligns both pixel-wise (MSE) and perceptual/semantic objectives (diagnosis-based severity reduction).
• Ablation shows removal of RL degrades both PSNR/SSIM and perceptual consistency, evidencing the necessity of RL-grounded alignment.

Results and Numerical Evaluation

The R&R framework was evaluated on OTS and RESIDE synthetic benchmarks, as well as on newly-collected real-world outdoor datasets, using peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) metrics.

Strong quantitative improvements over baselines:

• On OTS: R&R achieves 19.564 dB / 0.6214 SSIM (exceeding all previous methods including Stable Diffusion 3 and Qwen-Image-Edit).
• On RESIDE: 17.0036 dB / 0.6188 SSIM, showing robust performance across scene types.

Qualitative analysis demonstrates that conventional, prompt-based, or agentic IR methods (e.g., RestoreAgent, FoundIR, Qwen-Image-Edit) often fail under severe or mixed degradation, producing hallucinated content or ill-posed restoration. R&R reliably preserves scene geometry, color consistency, and semantic integrity.

Ablation studies underscore component necessity:

• Removing severity scoring, parameters, semantics, or RL alignment all reduce performance, each by a nontrivial margin.
• The synergy of structured reasoning and RL alignment yields optimal fidelity.

Implications and Future Directions

The R&R framework marks an architectural shift in image restoration: introducing holistic diagnostic reasoning tightly entangled with generative restoration, rather than relegating reasoning to either controller logic (as in agentic IR) or prompt design. This paradigm deepens restoration interpretability by offering detailed insight into model decisions and failures.

Practical implications:

• Improved robustness on real-world, adverse-environment images (e.g., autonomous driving, aerial inspection) where mixed degradations are the norm.
• Automatic, explicit quantification of degradation factors supports user trust and enables targeted post-processing or quality control pipelines.
• The modularity of CoT outputs enables transfer to other restoration/augmentation contexts and facilitates explainable AI in vision.

Theoretical and methodological impact:

• Establishes CoT reasoning as an actionable, trainable prior for generative models in vision tasks.
• Demonstrates the viability of RL-aligned multimodal pipelines, with supervision from diagnostically meaningful reward signals.

Possible future avenues:

• Expansion of the diagnostic schema to additional degradation types (beyond atmospheric and sensor noise) including adversarial corruptions or domain shifts.
• Incorporation of temporal reasoning in video restoration, where spatiotemporal semantics and degradation evolution can be leveraged.
• Exploration of unsupervised or few-shot diagnostic reasoning for domain adaptation and low-resource settings.

Conclusion

The Reason and Restore framework (2604.09511) exemplifies a methodologically rigorous, interpretable approach to UIR. By structurally decomposing the restoration process into explicit diagnostic reasoning and tightly aligning pixel reconstruction with semantic priors, R&R sets a new state of the art in both restoration quality and process transparency. This work strongly motivates the integration of reasoning-centric principles in future vision-language restoration pipelines and underscores the utility of reinforcement learning for semantic alignment in generative systems.