[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-85939-en":3,"doc-seo-85939-105":29,"detail-sidebar-cat-0-en-105":91},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":20,"is_deleted":4,"is_public":20,"is_downloadable":20,"audit_status":20,"page_count":21,"language":22,"language_code":23,"site_id":24,"html_lang":23,"table_of_contents":25,"faqs":26,"seo_title":13,"seo_description":14,"update_tm":27,"read_time":28},85939,7971461740909,"Levi","https://ap-avatar.wpscdn.com/davatar_155a257f0dc6eb9ab79c44ca47cae57d",6,"Technology","NanoVSR: Towards Real-Time Video Super-Resolution on Edge Devices","Recent Video Super-Resolution (VSR) systems often depend on transformer designs and explicit optical flow, which increase computational load and require custom operations that complicate deployment on edge-oriented hardware accelerators such as TensorRT. NanoVSR presents a scalable, fully convolutional architecture for resource-constrained devices. Through structural reparameterization, it collapses into standard convolutions at inference, preserving hardware compatibility. Progressive two-stage training enables implicit spatio-temporal alignment without motion compensation, achieving strong accuracy–efficiency trade-offs on REDS4.","arXiv :2607 . 10495v1 [ cs .CV] 11 Jul 2026  \nNanoVSR: Towards Real-Time Video Super-Resolution on Edge Devices  \nFilip Pawlicki 1, Marcel Kańduła2, Marcin Pucek 1, and Kamil Dobies 1  \n1 Gdańsk University of Technology, Faculty of Electronics, Telecommunications and  \nInformatics, Department of Computer Architecture, Gdańsk, Poland {s198371, s197893, [s197875}@student.pg.edu.pl](s197875}@student.pg.edu.pl)  \n2 Gdańsk University of Technology, Faculty of Electronics, Telecommunications and  \nInformatics, Department of Software Engineering, Gdańsk, Poland [s197677@student.pg.edu.pl](s197677@student.pg.edu.pl)  \nAbstract. Recent Video Super-Resolution (VSR) methods rely heavily on transformers and explicit optical flow, creating computational overhead and custom operations that hinder deployment on hardware accelerators like TensorRT. To address this, we introduce NanoVSR, a scalable, fully convolutional architecture designed for resource-constrained edge devices. Using structural reparameterization, NanoVSR collapses into standard convolutions during inference, ensuring seamless hardware compatibility and negligible runtime overhead. Furthermore, despite lacking explicit motion compensation, it maintains competitive restoration quality by implicitly learning spatio-temporal alignments through progressive training. Evaluated on the REDS4 benchmark, NanoVSR demonstrates an exceptional balance between accuracy and computational efficiency, significantly improving the trade-off for compact architectures.  \nOur NanoVSR-644k baseline yields 28.64 dB PSNR while delivering  \n27.2 FPS on the NVIDIA Jetson Orin NX 16GB (25W), offering massive speed gains over heavier models. The scaled NanoVSR-1.7M variant reaches 29.15 dB with a throughput of 19.58 FPS, providing superior, edge-optimized upscaling. Code is available at [https://github.com/](https://github.com/)[ ](https://github.com/)[filippawlicki/nanovsr](filippawlicki/nanovsr.)[.](filippawlicki/nanovsr.)  \nKeywords: Video Super-Resolution · Edge AI · Structural Reparameterization  \n1 Introduction  \nSuper-resolution (SR) is a fundamental challenge in computer vision aiming toreconstruct high-resolution (HR) images from low-resolution (LR) counterparts. While single image super-resolution focuses on recovering missing high-frequency details from a static image [8, 43], it is inherently limited by the absence of external information. Video super-resolution (VSR) transcends this limitation by introducing the temporal dimension, leveraging the rich, redundant information scattered across consecutive frames to recover details that are physically  \n2 F. Pawlicki et al.  \nabsent in a single frame. The fundamental mechanism of VSR involves the effective aggregation of this temporal data; however, because objects and cameras move, simply stacking neighboring frames results in ghosting and blurring. Consequently, the core challenge of VSR lies in finding effective methods for motion compensation.  \nRecent studies [3, 18, 38] have produced architectures capable of achieving impressive visual quality, successfully recovering details and maintaining temporal consistency across frames. To achieve this high standard of restoration, contemporary methods typically employ very deep neural networks that process video sequences using bidirectional propagation. Crucially, they rely on sophisticated alignment modules, utilizing dense optical flow estimation or complex attention mechanisms to maximize information extraction from multiple framesand ensure correct alignment before reconstruction.  \nHowever, achieving this impressive visual quality imposes a heavy demand on system resources due to the reliance on explicit motion compensation and deep feature extraction. While manageable on high-end GPUs, these architectures require memory and processing power that far exceed the capabilities of typical edge devices. Consequently, when deployed on computationally constrained platforms, the execution overhead ex","cbCaifSNhRgjfWFE","https://ap.wps.com/l/cbCaifSNhRgjfWFE","pdf",8639187,1,18,"English","en",105,"# Introduction\n## Background: Video vs. Single-Image Super-Resolution\n## Challenge: Motion Compensation and Resource Demands\n## Proposed Approach: NanoVSR","[{\"question\":\"Why do recent VSR methods struggle to run in real time on edge devices?\",\"answer\":\"They typically rely on explicit motion compensation such as dense optical flow and deep feature alignment modules. This increases memory use and compute overhead beyond what typical edge platforms can handle, preventing real-time deployment.\"},{\"question\":\"How does NanoVSR improve hardware compatibility during inference?\",\"answer\":\"NanoVSR uses structural reparameterization so the multi-branch convolutional structure can mathematically collapse into standard convolutions at inference. This avoids custom operations and reduces runtime overhead on accelerators.\"},{\"question\":\"How can NanoVSR maintain quality without explicit motion compensation?\",\"answer\":\"Despite discarding optical flow, NanoVSR trains a bidirectional recurrent framework with progressive two-stage training. The model implicitly learns spatio-temporal alignment and long-term dependencies, preserving competitive restoration quality.\"}]",1784207279,45,{"code":4,"msg":30,"data":31},"ok",{"site_id":24,"language":23,"slug":32,"title":13,"keywords":33,"description":14,"schema_data":34,"social_meta":86,"head_meta":88,"extra_data":90,"updated_unix":27},"nanovsr-towards-real-time-video-super-resolution-on-edge-devices","",{"@graph":35,"@context":85},[36,53,68],{"@type":37,"itemListElement":38},"BreadcrumbList",[39,43,47,50],{"item":40,"name":41,"@type":42,"position":20},"https://docshare.wps.com","Home","ListItem",{"item":44,"name":45,"@type":42,"position":46},"https://docshare.wps.com/document/","Document",2,{"item":48,"name":12,"@type":42,"position":49},"https://docshare.wps.com/document/technology/",3,{"item":51,"name":13,"@type":42,"position":52},"https://docshare.wps.com/document/nanovsr-towards-real-time-video-super-resolution-on-edge-devices/85939/",4,{"url":51,"name":13,"@type":54,"author":55,"headline":13,"publisher":57,"fileFormat":60,"inLanguage":23,"description":14,"dateModified":61,"datePublished":62,"encodingFormat":60,"isAccessibleForFree":63,"interactionStatistic":64},"DigitalDocument",{"name":9,"@type":56},"Person",{"url":40,"name":58,"@type":59},"DocShare","Organization","application/pdf","2026-07-17","2026-07-16",true,{"@type":65,"interactionType":66,"userInteractionCount":20},"InteractionCounter",{"@type":67},"ViewAction",{"@type":69,"mainEntity":70},"FAQPage",[71,77,81],{"name":72,"@type":73,"acceptedAnswer":74},"Why do recent VSR methods struggle to run in real time on edge devices?","Question",{"text":75,"@type":76},"They typically rely on explicit motion compensation such as dense optical flow and deep feature alignment modules. This increases memory use and compute overhead beyond what typical edge platforms can handle, preventing real-time deployment.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"How does NanoVSR improve hardware compatibility during inference?",{"text":80,"@type":76},"NanoVSR uses structural reparameterization so the multi-branch convolutional structure can mathematically collapse into standard convolutions at inference. This avoids custom operations and reduces runtime overhead on accelerators.",{"name":82,"@type":73,"acceptedAnswer":83},"How can NanoVSR maintain quality without explicit motion compensation?",{"text":84,"@type":76},"Despite discarding optical flow, NanoVSR trains a bidirectional recurrent framework with progressive two-stage training. The model implicitly learns spatio-temporal alignment and long-term dependencies, preserving competitive restoration quality.","https://schema.org",{"og:url":51,"og:type":87,"og:title":13,"og:site_name":58,"og:description":14},"article",{"robots":89,"canonical":51},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":92},[93,97,101,105,110,113,118,123,128,131,135],{"id":20,"doc_module":4,"doc_module_name":45,"category_name":94,"show_sort_weight":95,"slug":96},"Story & Novel",90,"story-novel",{"id":46,"doc_module":4,"doc_module_name":45,"category_name":98,"show_sort_weight":99,"slug":100},"Literature",80,"literature",{"id":52,"doc_module":4,"doc_module_name":45,"category_name":102,"show_sort_weight":103,"slug":104},"Exam",70,"exam",{"id":106,"doc_module":4,"doc_module_name":45,"category_name":107,"show_sort_weight":108,"slug":109},5,"Comic",60,"comic",{"id":11,"doc_module":4,"doc_module_name":45,"category_name":12,"show_sort_weight":111,"slug":112},50,"technology",{"id":114,"doc_module":4,"doc_module_name":45,"category_name":115,"show_sort_weight":116,"slug":117},7,"Healthcare",40,"healthcare",{"id":119,"doc_module":4,"doc_module_name":45,"category_name":120,"show_sort_weight":121,"slug":122},8,"Research & Report",30,"research-report",{"id":124,"doc_module":4,"doc_module_name":45,"category_name":125,"show_sort_weight":126,"slug":127},9,"Religion & Spirituality",20,"religion-spirituality",{"id":126,"doc_module":4,"doc_module_name":45,"category_name":129,"show_sort_weight":126,"slug":130},"World Cup","world-cup",{"id":132,"doc_module":4,"doc_module_name":45,"category_name":133,"show_sort_weight":132,"slug":134},10,"Lifestyle","lifestyle",{"id":136,"doc_module":4,"doc_module_name":45,"category_name":137,"show_sort_weight":106,"slug":138},19,"General","general"]