Fine-tuning text_encoder along with unet can bring us much more fidelity in the results when training Dreambooth. With LoRA, my personal experiments have shown that this conclusion holds true. Henc...