Why we need two tokenizer in the train_dreambooth_lora_sdxl

[daiyixiang666]

I am a starter and I have a question that why we need two tokenizer to trian the lora in train_dreambooth_lora_sdxl, I am some of confused

[daiyixiang666]

We want to get the pooled embedding, I got this

[sayakpaul]

SDXL uses two different text encoders. That's why.