RenderSR: A Lightweight Super-Resolution Model for Mobile Gaming Upscaling

Tingxing (Tim) Dong, Hao Yan, Mayank Parasar, Raun Krisch; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2022, pp. 3087-3095

Mobile game play can be a prime use case where an efficient SR network can lead to both performance boosts and power savings. In this paper, we present RenderSR (RSR), a bandwidth aware super-resolution network designed for use in mobile game upscaling. We explore how different factors affect the resulting image quality: color space, the inclusion of the depth channel, sharpening. With a 40K parameter size, RenderSR without sharpening achieves a PSNR value difference ranging -0.41 to 0.36dB from several much larger SR models. RenderSR with sharpening super resolved large objects such as rocks, buildings, tree trunks are almost identical to the ground truth. Based on our performance experiment, we propose that RenderSR upscales the GPU rendered image on NPU or DSP on the mobile SoC.