But by default, training algos don't see the lower resolution, your gradient just doesn't work as well. There is a body of research on how to make training awar

But by default, training algos don't see the lower resolution, your gradient just doesn't work as well. There is a body of research on how to make training aware of / adapt to the lower precision.