💻 computer science

PaAgent: Portrait-Aware Image Restoration Agent via Subjective-Objective Reinforcement Learning

The paper introduces PaAgent, a portrait-aware image restoration agent that utilizes a self-evolving portrait bank with Retrieval-Augmented Generation to efficiently select restoration tools and employs a subjective-objective reinforcement learning strategy to enhance degradation perception in complex scenes.

Original authors: Yijian Wang, Qingsen Yan, Jiantao Zhou, Duwei Dai, Wei Dong

Published 2026-03-19

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yijian Wang, Qingsen Yan, Jiantao Zhou, Duwei Dai, Wei Dong

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a very old, damaged photograph. It's blurry, covered in rain streaks, and the colors are faded. You want to fix it, but you don't know which specific tool to use first. Do you start by removing the rain? Or maybe you should brighten the dark spots first? If you pick the wrong tool, you might make the picture worse.

This is the problem the researchers faced with Image Restoration (IR). They wanted to build a computer "agent" (a smart software assistant) that could look at a bad photo, figure out what's wrong, and pick the perfect tool to fix it.

Here is how their new solution, PaAgent, works, explained simply:

1. The Problem: The "Exhaustive Search" Trap

Imagine you are in a giant library with thousands of repair manuals. If you have a broken photo, a standard computer agent might try to read every single manual one by one to see which one fits. It's slow, inefficient, and it often gets confused.

Previous agents were like this: they didn't remember what they had done before. Every time they saw a new photo, they started from scratch, guessing blindly. This led to "error accumulation"—trying the wrong fix, making a mess, and then trying to fix the mess, often failing completely.

2. The Solution: The "Smart Librarian" (The Portrait Bank)

The authors created PaAgent, which is like hiring a super-smart librarian who has read every repair manual in the world and remembers every single repair job they've ever done.

The Portrait Bank: Instead of guessing, PaAgent keeps a massive "memory bank" (called a Portrait Bank). Every time it fixes a photo, it writes down a "portrait" of that experience: What was wrong? What tool did we use? Did it work?
The RAG (Retrieval-Augmented Generation): When a new bad photo arrives, PaAgent doesn't guess. It quickly flips through its memory bank to find similar past cases. It asks, "Hey, I've seen a photo like this before! We used Tool X and it worked great."
The Result: It skips the trial-and-error phase. It goes straight to the right tool, saving time and avoiding mistakes.

3. The Brain: "Subjective-Objective" Reinforcement Learning

Even with a memory bank, the agent needs to learn how to judge quality. A computer can measure pixel sharpness (Objective), but it can't always "feel" if an image looks natural or if the colors are weird (Subjective).

The researchers taught PaAgent using a special training method called Subjective-Objective Reinforcement Learning (SORL). Think of this as a two-teacher system:

Teacher A (The Objective Robot): Uses math to measure quality. "This image has 90% sharpness. Good."
Teacher B (The Subjective Artist): A super-smart AI that looks at the image and says, "Hmm, the sky looks a bit too blue, and the shadows look unnatural. It needs more work."

PaAgent gets a "reward" only when both teachers are happy. If the math says it's good, but the artist says it looks weird, the agent learns that it made a mistake. This helps the agent understand not just if an image is clear, but how the damage is distributed (e.g., is the blur only in the corners, or everywhere?).

4. The Process: A Dynamic Conversation, Not a Script

Old methods were like following a rigid recipe: Step 1: Remove rain. Step 2: Fix blur. If Step 1 didn't work perfectly, Step 2 would still happen, ruining the photo.

PaAgent is more like a detective solving a mystery:

Perceive: "I see rain and low light."
Act: "Let's try removing the rain first."
Reflect: "Okay, the rain is gone, but the photo is still dark and hazy. My first guess was only half right."
Adapt: "Okay, I need to try a second tool to fix the haze."

It keeps checking its work after every step. If the photo is fixed, it stops. If not, it keeps going until the job is done right.

Summary

PaAgent is a smart image repair bot that:

Remembers past fixes so it doesn't have to guess (The Portrait Bank).
Learns from both math scores and human-like visual judgment (Subjective-Objective Learning).
Adapts its plan on the fly, checking its work after every step to ensure the photo is truly perfect.

The result? It fixes complex, messy photos (like a snowy, rainy, dark night scene) much better and faster than previous methods, acting less like a rigid machine and more like a skilled human restorer.

1. Problem Statement

Image Restoration (IR) aims to recover high-quality images from degraded ones (e.g., noise, haze, rain, low-light). While traditional All-in-One (AiO) methods attempt to handle multiple degradations within a single model, they often suffer from optimization conflicts, leading to a trade-off between generalization and restoration accuracy.

Recent Agent-based IR methods leverage Multi-modal Large Language Models (MLLMs) to perceive degradation and select restoration tools. However, existing agents face two critical limitations:

Lack of Insight Summarization: They treat each restoration instance in isolation, lacking a mechanism to summarize past interactions. This forces agents into exhaustive, inefficient searches for the optimal tool, even for simple degradations.
Ambiguous Degradation Perception: Existing agents often rely on a single No-Reference Image Quality Assessment (NR-IQA) score as a reward. A single score fails to capture the spatial distribution of degradations (e.g., distinguishing between local severe degradation and global minor degradation), leading to erroneous decision-making.

2. Methodology: PaAgent

The authors propose PaAgent, a portrait-aware IR agent designed to overcome these limitations through two core innovations: a Tool Portrait Bank and a Subjective-Objective Reinforcement Learning (SORL) strategy.

A. Tool Portrait Bank & Retrieval-Augmented Generation (RAG)

To avoid exhaustive tool searching, PaAgent maintains a dynamic Tool Portrait Bank that stores historical interaction insights.

Structure: Each entry is a triplet consisting of: (1) Degradation information, (2) The applied restoration tool, and (3) A subjective evaluation of the result.
Evolution: The bank is self-evolving. After a restoration task, the system (using Qwen3.5-Plus) summarizes the interaction (degraded image, restored image, selected tool) and stores the new insight.
RAG Mechanism: When a new degraded image is input, the agent encodes the degradation context and queries the vector database. It retrieves the top- $k$ relevant historical insights (chunks) and feeds them into the LLM (Qwen3.5-Plus) via prompt templates. This allows the agent to select the most suitable IR tool based on past successful experiences rather than random exploration.

B. Subjective-Objective Reinforcement Learning (SORL)

To enhance the agent's ability to perceive complex and partial degradations, the authors introduce SORL, which combines semantic understanding with quantitative metrics.

Reward Generation: Instead of relying solely on a single score, the reward generator (Qwen3.5-Plus) synthesizes:
1. Subjective Insights: Semantic analysis of the degradation and restoration quality by the MLLM.
2. Objective Scores: A composite of multiple NR-IQA metrics (CLIP-IQA, Hyper-IQA, NIQE, etc.).
Training Algorithm (GRPO): The paper employs Group Relative Policy Optimization (GRPO).
- For a given task, the agent generates a group of $n$ output trajectories.
- The reward generator assigns binary rewards ($0$ or $1$) to each trajectory based on the dual-perspective analysis.
- The policy is updated by calculating the normalized advantage of each trajectory relative to the group mean. This allows the model to learn from relative comparisons, refining its policy to select tasks that satisfy both visual preferences and objective quality standards.

C. Multi-Step Decision Workflow

Unlike rigid one-step decision agents, PaAgent operates in a perceive-restore-evaluate loop.

Perception: Analyze the image and retrieve relevant insights from the Portrait Bank.
Planning: Select the next restoration tool.
Execution: Apply the tool.
Reflection: Evaluate the intermediate result. If unsatisfactory, the loop continues; if satisfactory, the process terminates. This prevents premature termination and handles composite degradations effectively.

3. Key Contributions

Insight-Driven Tool Invocation: PaAgent is the first work to integrate a Tool Portrait Bank with RAG for image restoration. It enables the agent to leverage historical interaction insights to make informed tool selections, significantly reducing redundant trials.
Subjective-Objective Reinforcement Learning (SORL): A novel training strategy that uses GRPO to align policy updates with both semantic insights (MLLM) and quantitative quality scores (NR-IQA). This enables fine-grained perception of spatial degradation patterns.
Self-Evolving Knowledge Base: The system continuously enriches its portrait bank with new interaction insights, allowing the agent to improve its proficiency over time without retraining from scratch.

4. Experimental Results

The authors evaluated PaAgent on 8 benchmark datasets covering 6 single-degradation and 8 mixed-degradation scenarios (including CDD-11, LOL-V1, GoPro, SOTS, etc.).

Quantitative Performance: PaAgent achieved State-of-the-Art (SOTA) results in both PSNR and SSIM across almost all datasets.
- Example: On the CDD-11 dataset (complex mixed degradations), PaAgent significantly outperformed leading AiO methods (e.g., DGSolver, AdaIR) and the previous agent-based method (AgenticIR).
- Example: In the desnowing task (CSD), PaAgent achieved a PSNR of 37.95, surpassing the next best method (AgenticIR at 34.06).
Qualitative Performance: Visual comparisons show PaAgent better preserves fine textures (e.g., fur, foliage, building details) and handles composite degradations (e.g., low-light + blur + rain) without introducing color casts or over-smoothing.
Ablation Studies:
- Removing the Portrait Bank (Random Selection) caused a massive drop in performance.
- Removing the SORL strategy (using only SFT) resulted in suboptimal perception of local degradations.
- Removing input context (NR-IQA scores or task history) led to incorrect tool selection and sequencing.
- The Multi-step decision mechanism was proven crucial for handling complex scenarios where a single step is insufficient.

5. Significance

This paper represents a significant shift in the Image Restoration paradigm from static, single-model approaches to dynamic, agent-based reasoning.

Efficiency: By using RAG to retrieve historical insights, PaAgent avoids the "trial-and-error" inefficiency of current agents.
Robustness: The SORL strategy solves the ambiguity of single-score evaluation, allowing the agent to distinguish between global and local degradations, which is critical for real-world applications.
Scalability: The self-evolving nature of the Portrait Bank suggests a path toward agents that continuously learn and adapt to new degradation patterns without requiring massive retraining cycles.

In summary, PaAgent demonstrates that combining retrieval-augmented reasoning with subjective-objective reinforcement learning creates a more intelligent, efficient, and accurate system for complex image restoration tasks.