A Reduced-Precision Network for Image Reconstruction

Manu Mathew Thomas, Karthik Vaidyanathan, Gabor Liktor, and Angus G. Forbes
To appear in Proc. SIGGRAPH ASIA, December 2020.

Neural networks are often quantized to use reduced-precision arithmetic, as it greatly improves their storage and computational costs. This approach is commonly used in applications like image classification and natural language processing, however, using a quantized network for the reconstruction of HDR images can lead to a significant loss in image quality. In this paper, we introduce QW-Net, a neural network for image reconstruction, where close to 95% of the computations can be implemented with 4-bit integers. This is achieved using a combination of two U-shaped networks that are specialized for different tasks, a feature extraction network based on the U-Net architecture, coupled to a filtering network that reconstructs the output image. The feature extraction network has more computational complexity but is more resilient to quantization errors. The filtering network, on the other hand, has significantly fewer computations but requires higher precision. Our network uses renderer-generated motion vectors to recurrently warp and accumulate previous frames, producing temporally stable results with significantly better quality than TAA, a widely used technique in current games.

Manu Mathew Thomas, UC Santa Cruz
Karthik Vaidyanathan, Intel Corporation
Gabor Liktor, Intel Corporation
Angus Forbes, UC Santa Cruz


Paper (Author's version - PDF)
Supplementary Materials (ZIP)
Code (Coming soon!)
Data (Coming soon!)


link to video


   title = {A Reduced-Precision Network for Image Reconstruction},
   author = {Manu Mathew Thomas and Karthik Vaidyanathan and Gabor Liktor and Angus G. Forbes},
   journal = {ACM Trans. Graph.},
   volume = {39},
   number = {6},
   articleno = {231},
   numpages = {12},
   year = {2020},
   doi = {10.1145/3355089.3356565}


We thank Epic Games, Inc. for providing Unreal Engine with demo scenes for training and testing, and members of the Intel Advanced Research and Technology group for discussions and insightful feedback. We also thank David Blythe and Charles Lingle for supporting this research. We also thank Jan Novak (novakj4@gmail.com) and Benedikt Bitterli (benedikt.bitterli@gmail.com) for the script used for the Interactive Viewer on this webpage.