OARS: Process-Aware Online Alignment for Generative Real-World Image Super-Resolution

2026-03-13

AI Summary

This work addresses the challenge in generative real-world image super-resolution of balancing perceptual quality and fidelity while aligning with human visual preferences under unknown degradations. To this end, we propose OARS, a novel framework that introduces the first multimodal large language model (MLLM)-based COMPASS reward mechanism. OARS incorporates an input-quality-adaptive strategy for perceptual-fidelity trade-offs and employs a three-stage fine-grained annotation pipeline coupled with an online alignment training paradigm. Through progressive reinforcement learning—from full-reference to no-reference settings—and shallow LoRA optimization, OARS achieves state-of-the-art performance on the Real-ISR benchmark. Extensive user studies and experiments demonstrate that OARS significantly enhances perceptual quality while preserving high fidelity.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Computer Vision: Generative Adversarial Networks (GANs) for Vision
• Natural Language Processing: Generation

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Economics, Online Markets and Human Computation: LLM based quality controls for crowd work

Abstract

Aligning generative real-world image super-resolution models with human visual preference is challenging due to the perception--fidelity trade-off and diverse, unknown degradations. Prior approaches rely on offline preference optimization and static metric aggregation, which are often non-interpretable and prone to pseudo-diversity under strong conditioning. We propose OARS, a process-aware online alignment framework built on COMPASS, a MLLM-based reward that evaluates the LR to SR transition by jointly modeling fidelity preservation and perceptual gain with an input-quality-adaptive trade-off. To train COMPASS, we curate COMPASS-20K spanning synthetic and real degradations, and introduce a three-stage perceptual annotation pipeline that yields calibrated, fine-grained training labels. Guided by COMPASS, OARS performs progressive online alignment from cold-start flow matching to full-reference and finally reference-free RL via shallow LoRA optimization for on-policy exploration. Extensive experiments and user studies demonstrate consistent perceptual improvements while maintaining fidelity, achieving state-of-the-art performance on Real-ISR benchmarks.

Problem

• image super-resolution
• human visual preference
• perception-fidelity trade-off
• real-world degradations
• generative modeling

Innovation

• process-aware online alignment
• generative image super-resolution
• perception-fidelity trade-off
• MLLM-based reward
• LoRA optimization