Metasurfaces have revolutionized planar optics due to their prominent ability in light field manipulation. Recently the incorporation of machine learning has further improved computational efficiency and reduced the reliance on professionals in designing various metasurfaces. However, the prevalent methods still suffer from configuration complexity and expensive training costs due to more than one model or combination of the rule-driven algorithm. This paper proposes a deep learning-based paradigm using only one deep learning model for the end-to-end design of versatile metasurfaces. The adopted deep-enhanced Resnet acts both as the surrogate of the electromagnetic simulator in forward spectrum prediction and as the path for backward gradient descent optimization of the meta-atom structures in the paralleled calculation. Without loss of generality, a polarization-multiplexing holographic and a polarization-independent vortex metasurface were designed by this paradigm and successfully demonstrated in the terahertz range. The extremely simplified framework presented here will not only propel the design and application of metasurfaces in terahertz communication and imaging fields, but its universality will also accelerate the research and development of subwavelength planar optics across various wavelengths through the AI-enhanced design for optical devices.