Traditional diffusion models mostly use U-Net as the denoising network, whereas DiT replaces U-Net with a Transformer and has achieved performance superior to U-Net on image generation tasks.