Network architectures of the encoder, the estimator, and the decoder used in our work. The kernel size, the padding type, the activation function, and the output shape for each layer are specified when necessary. k and s refer to the kernel size and the output shape, respectively. In the encoder, c stands for the total number of channels after concatenating multi-band images and additional input (c = 6 for the SDSS dataset and the CFHTLS dataset; c = 10 for the KiDS dataset). The encoder outputs the latent vector υ_A and the vector υ_B with parallel fully connected layers. In the decoder, b stands for the number of optical bands for images (b = 5 for the SDSS dataset and the CFHTLS dataset; b = 4 for the KiDS dataset). The Leaky ReLU activation has a leaky ratio of 0.2. In the estimator, m stands for the number of redshift bins in the output (quoted in Table 1).