AnySR: Realizing Image Super-Resolution as Any-Scale, Any-Resource

TL;DR

This work introduces AnySR, a framework that converts existing arbitrary-scale SR methods into an effective any-scale, any-resource approach. By partitioning scale sets into groups and sharing parameters across a hierarchy of subnets, AnySR enables smaller-scale SR tasks to run with reduced compute while maintaining performance, and growing subnets to handle larger scales with minimal overhead. A key contribution is the any-scale enhancement via feature-interweaving, which injects scale information into features at regular intervals, preserving scale-aware processing and improving reconstruction quality. Empirical results on standard datasets show that AnySR achieves competitive PSNR/LPIPS with substantial FLOPs reductions, demonstrating practical value for deployment on devices with varying resources. The work also provides thorough ablations, complexity analyses, and out-of-distribution evaluations, underscoring AnySR’s generality and efficiency across backbones such as EDSR and RDN.

Abstract

In an effort to improve the efficiency and scalability of single-image super-resolution (SISR) applications, we introduce AnySR, to rebuild existing arbitrary-scale SR methods into any-scale, any-resource implementation. As a contrast to off-the-shelf methods that solve SR tasks across various scales with the same computing costs, our AnySR innovates in: 1) building arbitrary-scale tasks as any-resource implementation, reducing resource requirements for smaller scales without additional parameters; 2) enhancing any-scale performance in a feature-interweaving fashion, inserting scale pairs into features at regular intervals and ensuring correct feature/scale processing. The efficacy of our AnySR is fully demonstrated by rebuilding most existing arbitrary-scale SISR methods and validating on five popular SISR test datasets. The results show that our AnySR implements SISR tasks in a computing-more-efficient fashion, and performs on par with existing arbitrary-scale SISR methods. For the first time, we realize SISR tasks as not only any-scale in literature, but also as any-resource. Code is available at https://github.com/CrispyFeSo4/AnySR.

Figures (8)

• Figure 1: PSNR performance v.s. computing cost measured as GFLOPs for existing methods of ECBSR zhang2021edge, VDSR kim2016accurate, RCAN Zhang_2018_ECCV and EDSR Lim_2017_CVPR_Workshops. For fairness, all reported data is borrowed from wang2022adaptive and the test set is DIV2K agustsson2017ntire.
• Figure 2: Overview of the proposed AnySR: Includes Shallow Feature Extraction, Deep Feature Extraction with AnySR Blocks, and Image Reconstruction. AnySR Block automatically selects sub-net $F_t$ based on task complexity for feature extraction and enhancement.
• Figure 3: Framework of our any-scale enhancement including (a) any-scale enhancement pipeline and (b) feature-interweaving illustration.
• Figure 4: PSNR(dB) comparisons across all scales on different datasets of arbitrary-scale SR model SRNO srnowei2023super, its AnySR variants (through different subnets) highlighted by $^\dag$, and AnySR-retrained version (through the largest network) denoted by $^\ddag$.
• Figure 5: Visualization of existing arbitrary-scale SR models LIIF chen2021learning, SRNO srnowei2023super, Meta-SR hu2019meta, ArbSR wang2021learning, their AnySR variants (through different subnets) highlighted by $^\dag$, and AnySR-retrained version (through the largest network) denoted as $^\ddag$.