People seek high-spec TVs for immersive viewing, but high-spec content is scarce. For example, Netflix streams only up to 4K, and most platforms cannot keep pace with advanced displays. To close this gap, we present the world’s first high-performance AI accelerator that enhances both spatial and temporal resolution, integrated into premium TVs and monitors to deliver a superior visual experience.

The key challenge is that high-resolution and high-frame-rate output reduces AI computation resource for each pixel, causing quality loss and ghost effect. We propose asymmetric convolution kernels to improve quality at the same model complexity, and in-kernel-sampling deformable convolution to further enhance results. With these techniques, our system supports up to 8K-UHD 60 FPS video.

The spatial and temporal resolution enhancement system can be integrated into commercial high-spec displays, highlighting its potential for next-generation display technology. With custom PCB and FPGA boards, we built a real-time demo platform that upscales Quarter HD 30 FPS video to 4K-UHD 60 FPS, enabling interactive and prototype demonstrations.