We propose a convolution-like linearization strategy that accelerates pre-trained diffusion transformers for ultra-resolution image generation.