Leveraging differentiable NAS and abstract genetic algorithms for optimizing on-mobile VSR performance

Download PDF
Open Webpage

Xuncheng Liu, Weizhan Zhang, Tieliang Gong, Caixia Yan, Rui Li

Published: 2025, Last Modified: 02 Feb 2026Mach. Learn. 2025EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: The proliferation of mobile video applications has for real-time Video Super-Resolution (VSR) technologies, necessitating architectures that simultaneously achieve superior fidelity and ultra-low latency. Existing Neural Architecture Search (NAS) methods achieve efficient performance within fixed topologies and feature dimensionalities, but inherently struggle to meet the flexible design demands of lightweight on-mobile architectures. When applied to on-mobile VSR, their insufficient exploration under real-time constraints results in suboptimal performance, exposing a critical gap in current approaches. To address these challenges, we propose a novel bilevel NAS framework designed to systematically explore diverse topological configurations and feature dimensionalities within an expanded search space. Our framework incorporates a collection of independent differentiable search spaces. At the upper level, we employ an Abstract Genetic Algorithm (AGA) to explore a broad spectrum of topological variations and feature dimensionality configurations. At the lower level, we introduce a latency-aware differentiable NAS method, which facilitates efficient fidelity-latency optimization within each differentiable search space. Additionally, we provide theoretical convergence guarantees for our bilevel NAS framework, ensuring the discovery of optimal architectures. Experimental validation demonstrates that our method achieves state-of-the-art performance, delivering a 5.78ms inference latency while significantly improving fidelity metrics across standard benchmarks. The proposed framework not only breaks the fidelity-latency trade-off but also demonstrates remarkable efficiency in early AGA generations, reducing architecture search iterations from 100 to several generations while maintaining superior results. Our implementation is available at https://github.com/liuxunchenglxc/Reducio.