(Conceptually, the number of pixels remains the same but the result of demosaicing is RGB pixels so what triples is the number of channels.)

I hear it's good to perform demosaicing, denoising and super-resolution in one step, so perhaps that's what's happening here?

EDIT: on the video (section "Log is Half-Baked"), they also mention the processing includes tone mapping, color adjustment and lens distortion correction.