Fix image generation to handle LoRA adapters - load base model first, then apply LoRA weights