Low-Rank Adaptation (LoRA) is a standard way to fine-tune image diffusion models for artistic styles, with hundreds of thousands of such models shared online. Typically, analyzing a LoRA's style means generating sample images and then running image-based features like CLIP. That process is computationally expensive and depends on prompt selection. Researchers from the University of Toronto and Adobe propose skipping image generation entirely and analyzing the LoRA weight matrices themselves.
Their method treats the LoRA weights as a feature representation of the adapted style. Because real-world LoRAs lack ground-truth labels, the authors designed a controlled experimental setup using SD1.5, the most popular base model in online communities, across fine-art paintings and architectural photos with diverse prompts. They also ran a small-scale SDXL experiment to verify generalization. The weight-space features proved robust to variations in fine-tuning seeds and the specific images used during training.
In clustering and retrieval tasks on unseen LoRAs, the weight-based approach preserved fine-grained style distinctions and outperformed image-based methods. The authors conclude that LoRA weights alone are effective style descriptors, offering a cheaper and more direct way to organize and search the large ecosystem of shared fine-tuned models.