The technique of Low-Rank Adaptation (LoRA) has been widely used in text-to-image models for rendering images with precision. However, integrating multiple LoRAs, especially for complex compositions, remains a challenge. This paper introduces methodologies like LoRA Switch and LoRA Composite to manage multiple LoRAs efficiently.

• LoRA Switch alternates between LoRAs during the denoising process.
• LoRA Composite takes all LoRAs into account for a unified composition.
• A new testbed, ComposLoRA, was developed with 480 composition sets.
• The study shows improved performance over existing baselines using GPT-4V evaluation.

In providing solutions for integrating multiple LoRAs in image generation, the paper makes a critical contribution to enhancing the realism and complexity of AI-generated images, promoting the convergence of art and AI technology.