Graphics & Rendering

Qualcomm received 1 patent describing a rendering pipeline that routes completed GPU frames to a dedicated neural processor for image enhancement, rather than handling that work on the GPU itself. The architecture keeps the GPU free to continue rendering at its normal pace while the NPU runs inference-based processing on the output, with both operating at the same specified frame rate to avoid introducing delays between the two stages.

Adobe's 1 patent covers a system that uses a machine learning model to upscale low-resolution polygon meshes in real time, then applies a layer of residual corrections as higher-quality data becomes available. The ML model is trained on the same mesh dataset it will later reconstruct, which allows it to produce domain-specific results rather than relying on general-purpose upscaling. This approach lets a coarse mesh render at improved fidelity immediately, before the full data has finished transmitting.

Nintendo filed 1 patent for a camera system designed for voxel-based games, where the perspective can shift between a close follow view behind the player character and a wide aerial view of the surrounding terrain. The transition between modes is driven by the character's position and gaze direction, and each mode uses its own level-of-detail rendering settings. To handle the occlusion problems that arise when pulling the camera far back over voxel terrain, the system applies dithering transparency to geometry that would otherwise block the player's view of the area below.