K-LoRA: Unlocking Training-Free Fusion of Any Subject and Style LoRAs

Figure 1. Visual illustrations. (a) demonstrates the superior generative performance of our proposed K-LoRA using FLUX [3], where the object reference is presented on the left, the style reference on the right, and the generated image is shown in the center. In contrast, (b) compares our method with existing state-of-the-art methods, B-LoRA [7] and ZipLoRA [27], which tend to lose style or content information due to alterations in the original weight matrix or underutilization of the network structure. Our approach enhances the information captured by each LoRA matrix, thereby achieving superior fusion effects without requiring additional training.