Depth-aware guidance with self-estimated depth representations of diffusion models.