Would it be possible to add BREAK to training?

[TeKett]

I saw someone post on reddit that they successfully added it to the training side, but did not say how or what the result was. And have not answered when i asked.

Is this a possible thing? Since i feel like it could help with training, like it helps with generating. But im not sure if its the same thing as just duplicating the text image pair with a different caption. Im not even sure what really happens when you go above 75 tokens, i would assume the same thing but my ample time spent on generating images with large prompts is telling me its not.

Alternatively, what would happen if you boosted the token amount to more then 75? It would also help with improving the quality i'd think, esp with the unwanted random cut of point between the chunks. Now that we have diffusers we should be able to just change the config however we want, but the code need to also reflect that and not expect a default config but read it from the model.