[westurner]

Perhaps this is moving the goalposts a bit, but e.g. depixelation does appear to defy such a presumed limit due to apparent information content? Perhaps it is that the network reading the sensor carries additional information associated with the lower-resolution or additional-fields' sensor data?

https://github.com/krantirk/Self-Supervised-photo :

> Given a low-resolution input image, PULSE searches the outputs of a generative model (here, StyleGAN) for high-resolution images that are perceptually realistic and downscale correctly.

Maybe no amount of feature engineering can actually add information?