A LoRA is Worth a Thousand Pictures
The study investigates whether Low-Rank Adaptation (LoRA) weights alone can effectively represent artistic styles without needing generated images or original training data. The researchers developed a method to construct style embeddings directly from these LoRA weights using Principal Component Analysis (PCA) and a calibration strategy. Results show that these LoRA weight embeddings generalize well, preserving fine-grained styles and outperforming methods based on image features like CLIP.