FuseSR+: Spatial-Temporal Fusion for High-Scale Super-Resolution

The workload of real-time rendering is steeply increasing as the demand for high resolution, high refresh rates, and high realism rises, overwhelming most graphics cards. To mitigate this problem, one of the most popular solutions is to render images at a lower resolution to reduce rendering overhead and then manage to accurately upsample the low-resolution (LR) rendered image to the high-resolution (HR) results, a.k.a. super-resolution techniques. Theoretically, reconstructing a higher-resolution image requires integrating additional information beyond the rendered low-resolution image. Most existing methods predominantly rely on temporal information from low-resolution sources, such as historical frames. However, the temporal information accumulated frame-by-frame is generally too sparse to recover the fine details for high-scale super-resolution. To break the limit, this paper proposes an efficient and effective neural super-resolution method that utilizes non-aligned spatial and temporal information to achieve stable large-scale super-resolution. We introduce low-cost historical and current high-resolution auxiliary G-buffers as additional cues to enhance high-frequency detail reconstruction, spatial coherence, and temporal stability. To efficiently align and fuse various features — including spatial and temporal, low- and high-resolution features — under runtime constraints, we propose several novel technical modules, specifically, an H-Net architecture, a corresponding pixel shuffling and unshuffling layers for aligning low- and high-resolution features in a compact network, and a lightweight temporal feature recurrent module for aggregating HR temporal features. These techniques significantly reduce rendering overhead without noticeable quality deterioration. Experiments demonstrate that our method is able to produce temporally consistent reconstructions in both 4 × 4 and the more challenging 8 × 8 large-scale upsampling tasks at 4K resolution with real-time performance, while delivering substantially improved visual quality, significant performance gains, and remarkable temporal stability compared to existing methods.

Authors

Institutions

Publication Details

Journal
ACM Transactions on Graphics
Published
2026-11-19
DOI
https://doi.org/10.1145/3859872
Primary Topic
Advanced Image Processing Techniques
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

FuseSR+: Spatial-Temporal Fusion for High-Scale Super-Resolution

Jingsen Zhu, Chuankun Zheng, Yuchi Huo, Weijia Xue et al.
ACM Transactions on Graphics
Advanced Image Processing Techniques
article

FuseSR+: Spatial-Temporal Fusion for High-Scale Super-Resolution

Jingsen Zhu, Chuankun Zheng, Yuchi Huo, Weijia Xue, Rui Wang, Shi Li, Tian Xie, Suiliang Mai
article en

Abstract

The workload of real-time rendering is steeply increasing as the demand for high resolution, high refresh rates, and high realism rises, overwhelming most graphics cards. To mitigate this problem, one of the most popular solutions is to render images at a lower resolution to reduce rendering overhead and then manage to accurately upsample the low-resolution (LR) rendered image to the high-resolution (HR) results, a.k.a. super-resolution techniques. Theoretically, reconstructing a higher-resolution image requires integrating additional information beyond the rendered low-resolution image. Most existing methods predominantly rely on temporal information from low-resolution sources, such as historical frames. However, the temporal information accumulated frame-by-frame is generally too sparse to recover the fine details for high-scale super-resolution. To break the limit, this paper proposes an efficient and effective neural super-resolution method that utilizes non-aligned spatial and temporal information to achieve stable large-scale super-resolution. We introduce low-cost historical and current high-resolution auxiliary G-buffers as additional cues to enhance high-frequency detail reconstruction, spatial coherence, and temporal stability. To efficiently align and fuse various features — including spatial and temporal, low- and high-resolution features — under runtime constraints, we propose several novel technical modules, specifically, an H-Net architecture, a corresponding pixel shuffling and unshuffling layers for aligning low- and high-resolution features in a compact network, and a lightweight temporal feature recurrent module for aggregating HR temporal features. These techniques significantly reduce rendering overhead without noticeable quality deterioration. Experiments demonstrate that our method is able to produce temporally consistent reconstructions in both 4 × 4 and the more challenging 8 × 8 large-scale upsampling tasks at 4K resolution with real-time performance, while delivering substantially improved visual quality, significant performance gains, and remarkable temporal stability compared to existing methods.

ACM Transactions on Graphics
Cornell University (US), Zhejiang University (CN)
Advanced Image Processing Techniques
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon