The current implementation uses a random latent vector rather than encoding your actual image. For more accurate results, a proper encoder would need to be implemented.