[Feat]: text "jitter"

[yggdrasil75]

Describe your use-case.

while I was writing my captions, I realized that I made a mistake on one image and had a characters name lowercased while in the rest of the training data it was correct. realizing that this will be more common in generation than in more detailed lora creation, I propose the following idea: text jitters. basically, similar to how there is the option for random rotations and random brightness, crop jitter, etc, similar parameters for text would be useful:

What would you like to see as a solution?

• Randomly Capitalize First Letters In A Tag
  • this will allow for those whos native keyboards autocorrect to capitals even when it doesnt make sense in english (I think that it was a german who complained about this happening, cant remember)
• randomly apply all lowercase
  • this helps for those of us who are lazy.
• randomly drop doubled leters
  • common spelling mistakes and all that.
• randomly mispeel words
  • at a lowered rate than above for less common mistakes
• random tag dropouts
  • this helps with overly long captions causing underfitting. ie: if I put "red hair" in every single caption for a character with red hair and the user doesnt, then the character may not have red hair in the results without it being intentional. dropping the tag "red hair" every once in a while would help with this
  • would also like this to have "goals" ie: drop random tags (other than the first n tags, similar to shuffle option) until down to 75 tokens, or until down to 150, or 225.

Have you considered alternatives? List them here.

an alternative would be to do this myself in the captions, but that wouldnt really be practical.