PocketSR: The Super-Resolution Expert in Your Pocket Mobiles

2025-10-03

AI Summary

To address the high computational cost and latency of real-world image super-resolution (RealSR) on edge devices, this paper proposes PocketSR, a lightweight single-step generative model. Our approach introduces three key innovations: (1) a LiteED encoder-decoder architecture with only 2.5% of the parameters of a conventional VAE; (2) an online annealing pruning strategy integrated into a U-Net backbone to dynamically compress redundant channels; and (3) a multi-level feature distillation loss to effectively transfer high-quality generative priors. With merely 146 million parameters, PocketSR processes 4K images in 0.8 seconds while achieving PSNR and SSIM scores competitive with state-of-the-art single-step and multi-step methods. By significantly reducing inference latency and memory footprint without compromising reconstruction fidelity, PocketSR enhances both inference efficiency and deployment feasibility on resource-constrained edge platforms.

Technology Category

• Machine Learning: Learning on the Edge & Model Compression
• Computer Vision: Diffusion Models for Vision
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Efficiency and scalability of Web search engines
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications
• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation

Abstract

Real-world image super-resolution (RealSR) aims to enhance the visual quality of in-the-wild images, such as those captured by mobile phones. While existing methods leveraging large generative models demonstrate impressive results, the high computational cost and latency make them impractical for edge deployment. In this paper, we introduce PocketSR, an ultra-lightweight, single-step model that brings generative modeling capabilities to RealSR while maintaining high fidelity. To achieve this, we design LiteED, a highly efficient alternative to the original computationally intensive VAE in SD, reducing parameters by 97.5% while preserving high-quality encoding and decoding. Additionally, we propose online annealing pruning for the U-Net, which progressively shifts generative priors from heavy modules to lightweight counterparts, ensuring effective knowledge transfer and further optimizing efficiency. To mitigate the loss of prior knowledge during pruning, we incorporate a multi-layer feature distillation loss. Through an in-depth analysis of each design component, we provide valuable insights for future research. PocketSR, with a model size of 146M parameters, processes 4K images in just 0.8 seconds, achieving a remarkable speedup over previous methods. Notably, it delivers performance on par with state-of-the-art single-step and even multi-step RealSR models, making it a highly practical solution for edge-device applications.

Problem

• Developing lightweight super-resolution models for mobile deployment
• Reducing computational costs while maintaining image quality
• Enabling efficient real-world image enhancement on edge devices

Innovation

• Ultra-lightweight single-step model for mobile super-resolution
• Efficient LiteED VAE alternative with 97.5% parameter reduction
• Online annealing pruning with multi-layer feature distillation