Foveated streaming is wild to me. Saccades are commonly as low as 20-30ms when reading text, so guaranteeing that latency over 2.4Ghz seems Sisyphean.
I wonder if they have an ML model doing partial upscaling until the eyetracking state is propagated and the full resolution image under the new fovea position is available. It also makes me wonder if there's some way to do neural compression of the peripheral vision optimized for a nice balance between peripheral vision and hints in the embedding to allow for nicer upscaling.
I wonder if they have an ML model doing partial upscaling until the eyetracking state is propagated and the full resolution image under the new fovea position is available. It also makes me wonder if there's some way to do neural compression of the peripheral vision optimized for a nice balance between peripheral vision and hints in the embedding to allow for nicer upscaling.