Rethinking LoRA approaches for normal photorealistic complex scenes with just SDXL base model. (more info in the comments)