fix: loading problem for sdxl lora dreambooth (huggingface#7166)

examples/dreambooth/train_dreambooth_lora_sdxl.py

@@ -1178,7 +1178,7 @@ def load_model_hook(models, input_dir):
             _set_state_dict_into_text_encoder(lora_state_dict, prefix="text_encoder.", text_encoder=text_encoder_one_)
 
             _set_state_dict_into_text_encoder(
-                lora_state_dict, prefix="text_encoder_2.", text_encoder=text_encoder_one_
+                lora_state_dict, prefix="text_encoder_2.", text_encoder=text_encoder_two_
             )
 
         # Make sure the trainable params are in float32. This is again needed since the base models