I know the text should be mirrored, but that would make it harder to read.

Base trigger: signing

Action 1: holding pen

Action 2: finger writing

There are other creative actions such as holding lipstick tube but they are not included in the training dataset.

Just holding pen without signing is also ok if you want to add the signing text in PS later.

To make precise and meaningful text generation, you need to make combined use of the LoRA with control net (I'm using the canny one: https://civarchive.com/models/929685?modelVersionId=1040650), here is the webui setting I'm using for the Miku example (also included in PNG INFO):

"invert" preprocessor is needed if you have black text and white background.

You may want to tweak the "Control Weight" and "Ending Control Step": lower control weight enables colorful text but it may break the letter; lower ending step makes the text more blended in and more natural but may also break the letter in later steps.