Let me tell you: the training data did not contain that many X to begin with :D
The Xs at best look like Ts, at worst like malformed Os.
Always funny to find the blind spots of some piece of tech.
[EDITED to add:] Ah, oops, I see. You were talking specifically about _capital_ X. You are right and I was wrong: the model produces disastrous capital Xs even at high legibility settings.
as usual, the errors are at least as interesting as the successes on new tech!
The same settings but with min speed got caught at the z, creating what looked like repeating waves then a few letters that I couldn't place from the phrase.
Also amusing are the loops that get the tops cropped even though other letters extend higher.
> My first name is: XXX ZZZ JJJ
(max speed, max legibility, default stroke)
To be fair, I kind of remember handwriting X (and x) as totally counter-intuitive in school.