Finetuning SDXL on long prompts

I have been trying to finetune SDXL with some dalle 3 style long descriptive prompt and I have been running into the weirdest issues (If it’s not compel somehow not working with one random long prompt, it’s some mismatch tensor issue that randomly pops up, or backward graph issue). I have been going at this task for like 3 days nonstop rn and wanna know if someone already did the work and Im just banging my head on the wall for nothing.
I have been using dreambooth lora sdxl as a base, but with some modification (added compel and some try catch blocks here and there). I can provide the clone/diff if needed.