This paper discusses the advancement of a state-of-the-art image generation model that enables personalized image generation with both user-defined subject matter (content) and style. Previous research achieved personalization by merging low-rank adapters (LoRA) using optimization-based methods, but this approach is computationally expensive and unsuitable for real-time use on resource-constrained devices such as smartphones. To address this issue, this paper proposes a LoRA$.$rar method that improves image quality while accelerating the merging process by over 4,000x. By pretraining a hypernetwork on diverse content-style LoRA pairs, we learn an efficient merging strategy that generalizes to new content-style pairs, enabling fast, high-quality personalization. Furthermore, we identify the limitations of existing content-style quality assessment metrics and propose a novel protocol that utilizes a multimodal large-scale language model (MLLM) for more accurate assessment. MLLM and human evaluations demonstrate that our method outperforms the state-of-the-art in both content and style fidelity.
