multi-concept

This paper proposes W-Switch and W-Composite, two training-free methods for combining multiple LoRA modules in text-to-image generation using prompt-aware importance weighting to reduce concept interference. The approach is evaluated on the ComposLoRA testbed and shows consistent improvements over state-of-the-art methods in visual quality and identity preservation.