Self-Cross Diffusion Guidance for Text-to-Image Synthesis of Similar Subjects

a rose and a carnation a leopard and a tiger a beagle and a collie a eagle and an owl SD1.4 [41] INITNO [17] CONFORM [31] Self-Cross (Ours) SD1.4 [41] INITNO [17] CONFORM [31] Self-Cross (Ours) Figure 1. Our Self-Cross guidance addresses subject mixing in particular for similar subjects. Our training-free method can boost the performance of any Unet-based or transformer-based diffusion models such as Stable Diffusion 1, 2, and 3 (shown in Fig. 5 and 6.