You should look at the raw (not diff)readings. Without any touch, you'll see the baseline values. Any difference across pads you see there will affect the clipping level of each pad. These differences are due to the geometry of the wiring and of the sensing pads. For instance, larger pads will have higher baseline values.
Ideally, you'd set the prescaler as high as necessary so that no pad clips when touched. Then revert to DIFF for normal use. You may still end up with different maximum DIFF values for each pad (e.g.: larger pads may give higher readings because of the larger contact surface with the finger). That's not necessarily bad, it just needs to be taken into account when you use the data.
If you have pads with widely different geometry, for instance some very small ones and some very large ones, you may need to adjust the prescaler to different values for each set of pads. That's slower and cumbersome, but doable.
See here for more context: https://learn.bela.io/using-trill/settings-and-sensitivity/
