Generating computer-generated holograms (CGHs) for 3D scenes by learning-based methods can reconstruct arbitrary 3D scenes with higher quality and faster speed. However, the homogenization and difficulty of obtaining 3D high-resolution datasets seriously limit the generalization ability of the model. A novel approach is proposed to train 3D encoding models based on convolutional neural networks (CNNs) using 2D image datasets. This technique produces virtual depth (VD) images with a statistically uniform distribution. This approach employs a CNN trained with the angular spectrum method (ASM) for calculating diffraction fields layer by layer. A fully convolutional neural network architecture for phase-only encoding, which is trained on the DIV2K-VD dataset. Experimental results validate its effectiveness by generating a 4K phase-only hologram within only 0.061 s, yielding high-quality holograms that have an average PSNR of 34.7 dB along with an SSIM of 0.836, offering high quality, economic and time efficiencies compared to traditional methods.