BREAK Keyword Trick for A1111 and LoRa Models in A1111 (Patreon)

This helped me crack the code on POV Missionary! The missionary Sex positions has alot of room for errors, like hands and their placement, the odd thighs, the art style(in the other POV one), and quite a few ones I had.

Cowgirl and squatting I didn't need to worry about the hand placements of the character I was doing since most of the training data had their hands out of view(squatting had them behind or in front below the intercourse, so I can easily crop them out), Missionary had alot of hands in view for training.

Taken from Reddit on the BREAK Keyword for A1111

"In the past there was actually a 75  token limit on prompts and it would just not process anything after 75  tokens. At some point they expanded this by separating longer prompts  into 75 token chunks. So now when your prompt goes over 75 tokens it  will start a new chunk, process each chunk separately, then add all the  chunks together.

BREAK serves as a way you can tell it to manually break the prompt so you can control how your prompt gets processed.

For  instance, your prompt goes over 75 tokens, at which point it would  automatically break it at the 75th token and start a new chunk, but  perhaps it's in the middle of a string of tokens that you want processed  together, so you would use BREAK to manually break the prompt where you  want it."

So if i had a large set of prompts at 190 tokens, every 75 tokens it will break the prompt up, when you manually break the prompt instead the prompts that you want together won't get broken up, having a series of prompts broken up messes with how the image generates and you will end up with inconsistent results because of the way the prompts were improperly broken up.

Prompt set 1, prompt set 2, prompt set 3, each set has 50 tokens, so 150 in total tokens. now it will break half of prompt set 2(25 tokens in first chunk and 25 in the second chunk)(that needed to be in the same chunk for image generation) so your images end up looking off and get strange and random results.

so now I do this,

Prompt set 1, BREAK, Prompt set 2, BREAK, Prompt set 3.

Now each set are in their own chunk and not split up.

New trick using BREAK for LoRa models as well!

When I used BREAK I had the character and Sex position model in the same chunk, I couldn't do dynamic LoRa due to Tyuule's ears not showing up so I had to use her full model weight to get her ears right.

So I thought after having issues on the layout of the images again with Tyuule, I realized what if I move the sex position LoRa to a chunk before the character one, and It may have solved my issue on her and for other characters that will have this problem.

And yes! It seems to have worked!

Something you should know is that prompts are used and prioritized in the order of top to bottom and left to right, but LoRa models play out no matter where they are placed, So by using the BREAK keyword It may have forced the LoRa model to follow the order based on the order of prompt chunks. I finally got a character showing up in the right layout thanks to this trick!

Example:

Chunk 1 (quality prompts)

BREAK

Chunk 2 (With sex lora) and related prompts, like nude, vaginal, sex, etc, etc.)

BREAK

Chunk 3(with character LoRa) and expressions with the related character prompts)