An autoencoder is a neural network that generates data highly similar to the input data for output. Although an autoencoder theoretically produces output almost identical to the input upon completion of learning, it actually generates blurred outputs for complex face images due to the omission of detailed information during the compression process and the use of MSE loss during learning. This paper addresses these issues by mapping detailed information from the frequency domain onto the latent space, adding to the existing latent vector, and learning using a mixed loss of MS-SSIM (Multi Scale Structural Similarity Index Measure) loss and l1 loss instead of MSE loss. As a result, the 100 x l1, 100 x l2 loss, SSIM, MS-SSIM between input and output are 12, 3.1, 0.53, and 0.575 respectively, leading to the production of images of higher quality than the standard autoencoder.