About LoRA model import

[liufeiluo]

Thanks to the authors for their advanced ideas. I have a trained LoRA model based on SDXL. I loaded that model with pipeline.load_lora_weights(lora_model_id) as I always do, but the result of getting large resolution is not good, is there something wrong with the model loaded this way?

[xtudbxk]

While I am not familiar with the lora code, I can give few suggestions.

For our FreCaS, we have three wrappers to adjust the behavior of scheduler, unet forward and attention maps.
Maybe you can check it by switching off the corresponding wrappers to see what happened.

Besides, another possible reason is that the optimal settings for the framework are channged. In such case, you can adjust the settings according to the ablation studies results.